封面设计：Ann Kirchner

Cover design by Ann Kirchner

封面图片 © pzAxe / Shutterstock.com

Cover image © pzAxe / Shutterstock.com

阿歇特图书集团支持言论自由权和版权价值。版权的目的是鼓励作家和艺术家创作出丰富我们文化的创意作品。

Hachette Book Group supports the right to free expression and the value of copyright. The purpose of copyright is to encourage writers and artists to produce the creative works that enrich our culture.

未经许可扫描、上传和分发本书是对作者知识产权的盗窃。如果您希望获得本书材料的使用许可（除用于评论目的外），请联系 permissions@hbgusa.com。感谢您对作者权利的支持。

The scanning, uploading, and distribution of this book without permission is a theft of the author’s intellectual property. If you would like permission to use material from the book (other than for review purposes), please contact permissions@hbgusa.com. Thank you for your support of the author’s rights.

基础书籍

Basic Books

阿歇特图书集团

Hachette Book Group

纽约州纽约市美洲大道 1290 号，邮编 10104

1290 Avenue of the Americas, New York, NY 10104

www.basicbooks.com

由 Basic Books 出版，是阿歇特图书集团 (Hachette Book Group, Inc.) 子公司 Perseus Books, LLC 的印记。Basic Books 名称和徽标是阿歇特图书集团的商标。

Published by Basic Books, an imprint of Perseus Books, LLC, a subsidiary of Hachette Book Group, Inc. The Basic Books name and logo is a trademark of the Hachette Book Group.

Hachette Speakers Bureau 为高峰活动提供大量作者。如需了解更多信息，请访问 www.hachettespeakersbureau.com 或致电 (866) 376-6591。

The Hachette Speakers Bureau provides a wide range of authors for peaking events. To find out more, go to www.hachettespeakersbureau.com or call (866) 376-6591.

出版商不对不属于其的网站（或其内容）负责。

The publisher is not responsible for websites (or their content) that are not owned by the publisher.

美国国会图书馆出版品目錄數據

Library of Congress Cataloging-in-Publication Data

姓名：福特，马丁（Martin R.）作者。

Names: Ford, Martin (Martin R.) author.

标题：机器人的统治：人工智能将如何改变一切/马丁·福特。

Title: Rule of the robots: how artificial intelligence will transform everything / Martin Ford.

描述：纽约：Basic Books，[2021] | 包括书目参考和索引。|

Description: New York: Basic Books, [2021] | Includes bibliographical references and index. |

标识符：LCCN 2021016340 | ISBN 9781541674738（精装本）| ISBN 9781541674721（电子书）

Identifiers: LCCN 2021016340 | ISBN 9781541674738 (hardcover) | ISBN 9781541674721 (ebook)

主题：LCSH：人工智能——社会方面。| 机器人技术——社会方面。

Subjects: LCSH: Artificial intelligence—Social aspects. | Robotics—Social aspects.

分类：LCC Q334.7 .F67 2021 | DDC 006.301—dc23

Classification: LCC Q334.7 .F67 2021 | DDC 006.301—dc23

LC 记录可在 https://lccn.loc.gov/2021016340 上查阅

LC record available at https://lccn.loc.gov/2021016340

ISBN：978-1-5416-7473-8（精装本）；978-1-5416-7472-1（电子书）

ISBNs: 978-1-5416-7473-8 (hardcover); 978-1-5416-7472-1 (ebook)

E3-20210813-JV-NF-ORI

献给我的母亲 Sheila

For my mother, Sheila

探索图书赠送、抢先体验、交易等活动。

Explore book giveaways, sneak peeks, deals, and more.

点击此处了解更多。

基础图书徽标

第一章

CHAPTER 1 新兴颠覆

THE EMERGING DISRUPTION

2020 年 11 月 30 日，谷歌母公司 Alphabet 旗下位于伦敦的人工智能公司 DEEPMIND 宣布了一项令人震惊且可能具有历史意义的计算生物学突破，这项创新有可能真正改变科学和医学。该公司成功地利用深度神经网络根据细胞中分子的遗传密码预测蛋白质分子将如何折叠成最终形状。这是一个里程碑，它结束了长达五十年的科学探索，标志着一项新技术的到来，这项技术将带来对生命结构的前所未有的理解，以及医学和制药创新的新时代^。1

ON NOVEMBER 30, 2020, DEEPMIND, A LONDON-BASED ARTIFICIAL intelligence company owned by Google parent Alphabet, announced a stunning, and likely historic, breakthrough in computational biology, an innovation with the potential to genuinely transform science and medicine. The company had succeeded in using deep neural networks to predict how a protein molecule will fold into its final shape based on the genetic code from which the molecule is constructed in cells. It was a milestone that culminated a fifty-year scientific quest and marked the advent of a new technology that was poised to usher in an unprecedented understanding of the very fabric of life—as well as a new age of medical and pharmaceutical innovation.¹

蛋白质分子是长链，每个链节由二十种不同的氨基酸组成。DNA 中编码的基因列出了组成蛋白质分子的氨基酸的精确序列，或者说是配方。然而，这种遗传配方并没有指定分子的形状，而这对其功能至关重要。相反，形状是由分子在细胞中制造后几毫秒内自动折叠成高度复杂的三维结构的方式决定的^。2

Protein molecules are long chains in which each link consists of one of twenty different amino acids. The genes encoded in DNA lay out the precise sequence, or essentially the recipe, of the amino acids that make up the protein molecule. This genetic recipe, however, does not specify the shape of the molecule, which is critical to its function. Instead, the shape results from the way the molecule automatically folds into a highly complex three-dimensional structure within milliseconds of its fabrication in the cell.²

预测蛋白质分子折叠的确切结构是科学界最艰巨的挑战之一。可能的形状数量几乎是无限的。科学家们将整个职业生涯都投入到这个问题中，但总体上只取得了有限的成功。DeepMind 的系统使用了该公司在 AlphaGo 和 AlphaZero 系统中开创的人工智能技术，这两个系统曾在围棋和国际象棋等棋盘游戏中击败了世界上最优秀的人类选手。但人工智能主要与游戏熟练度相关的时代显然即将结束。AlphaFold 能够预测蛋白质分子的形状，其准确度可与使用 X 射线晶体学等技术进行昂贵且耗时的实验室测量相媲美，这提供了无可辩驳的证据，表明人工智能最前沿的研究已经产生了一种实用且不可或缺的科学工具，具有改变世界的潜力。

Predicting the exact configuration into which a protein molecule will fold is one of the most daunting challenges in science. The number of possible shapes is virtually infinite. Scientists have devoted entire careers to the problem, but have collectively achieved only modest success. DeepMind’s system uses AI techniques that the company pioneered in the AlphaGo and AlphaZero systems that had famously triumphed over the world’s best human competitors at board games like Go and chess. But the era of AI being primarily associated with adeptness at games is clearly drawing to a close. AlphaFold’s ability to predict the shape of protein molecules with an accuracy that rivals expensive and time-consuming laboratory measurement using techniques like X-ray crystallography offers irrefutable evidence that research at the very frontier of artificial intelligence has produced a practical and indispensable scientific tool with the potential to transform the world.

当时地球上几乎每个人都可能已经见过最著名的蛋白质分子三维形状如何定义其功能的例子——冠状病毒刺突蛋白，一种允许病毒附着并感染宿主的分子对接机制——这一突破为我们带来了希望，让我们能够更好地为下一次大流行做好准备。该系统的一个重要用途可能是快速筛选现有药物，以找到最有可能对抗新出现的病毒的药物，从而在疫情爆发的最早阶段将强有力的治疗方法交到医生手中。除此之外，DeepMind 的技术有望带来各种进步，包括设计全新的药物和更好地了解蛋白质错误折叠的方式——这与糖尿病、阿尔茨海默氏症和帕金森氏症等疾病有关。这项技术未来可能会在医学以外的一系列应用中得到应用，例如，帮助设计能够分泌能够降解塑料或石油等废物的蛋白质的微生物。³换句话说，这是一项有可能加速几乎所有生物化学科学和医学领域进步的创新。

Arriving at a moment when nearly everyone on earth had likely encountered illustrations featuring the most notorious example of how a protein molecule’s three-dimensional shape defines its function—the coronavirus spike protein, a kind of molecular docking mechanism that allows the virus to attach to and infect its host—the breakthrough offered hope that we will be far better prepared for the next pandemic. One important use of the system might be to rapidly screen existing medications to find the ones likely to be most effective against a newly emergent virus, putting powerful treatments in the hands of doctors in the earliest stages of an outbreak. Beyond this, DeepMind’s technology is poised to lead to a variety of advances, including the design of entirely new drugs and a better understanding of the ways in which proteins can misfold—something that has been associated with illnesses like diabetes as well as Alzheimer’s and Parkinson’s diseases. The technology might someday be employed in a range of applications outside medicine, for example, to help engineer microbes that can secrete proteins capable of degrading waste such as plastic or oil.³ In other words, it is an innovation with the potential to accelerate progress in virtually every sphere of biochemical science and medicine.

在过去的十年里，人工智能领域取得了革命性的飞跃，并开始提供越来越多的实际应用，这些应用已经改变了我们周围的世界。这一进步的主要推动力是“深度学习”——一种基于 DeepMind 所采用的多层人工神经网络的机器学习技术。几十年来，人们已经了解了深度神经网络的基本原理，但最近的重大进展是由信息技术领域两大趋势的融合推动的：首先，功能强大得多的计算机的出现首次使神经网络转变为真正有能力的工具。其次，现在整个信息经济中生成和收集的海量数据为训练这些网络执行有用任务提供了至关重要的资源。事实上，曾经难以想象的大规模数据可用性可以说是我们所看到的惊人进步背后最重要的因素。深度神经网络吸收和利用数据的方式类似于巨大的蓝鲸以微小的磷虾为食，吞噬大量单个微小的生物，然后利用它们的集体能量来塑造一个体型巨大、力量强大的生物。

Over roughly the past decade, the field of artificial intelligence has taken a revolutionary leap forward and is beginning to deliver an ever-increasing number of practical applications that are already transforming the world around us. The primary accelerant of this progress has been “deep learning”—a machine learning technique based on the use of multilayered artificial neural networks of the kind employed by DeepMind. The basic principles of deep neural networks have been understood for decades, but recent dramatic advances have been enabled by the confluence of two relentless trends in information technology: First, the arrival of vastly more powerful computers has, for the first time, allowed neural networks to transition into truly capable tools. And, second, the enormous troves of data now being generated and collected across the information economy provide a resource crucial to training these networks to perform useful tasks. Indeed, the availability of data at a scale that would have once been unimaginable is arguably the single most important factor underlying the startling progress we have seen. Deep neural networks hoover up and leverage data much in the way that a massive blue whale feeds on tiny krill, scooping up vast numbers of individually insignificant organisms and then using their collective energy to animate a creature of magnificent size and power.

随着人工智能成功应用于越来越多的领域，很明显它正在演变成一项独一无二的重大技术。例如，在医学的某些特定领域，诊断人工智能应用已经开始匹敌甚至超越最优秀的医生。这种创新的真正力量不仅在于它有可能超越一位世界级医生，还在于技术中蕴含的智能可以轻松扩展。不久的将来，精英诊断专业知识将以可承受的价格传播到全球，即使在人们几乎无法接触任何医生或护士的地区也能获得它——更不用说世界上最好的医学专家了。

As artificial intelligence is successfully applied to more and more areas, it is becoming clear that it is evolving into a uniquely consequential technology. In some specific areas of medicine, for example, diagnostic AI applications are already beginning to match or even exceed the performance of the best doctors. The true power of such an innovation does not lie just in its ability to potentially outperform a single world-class physician, but rather in the ease with which the intelligence encapsulated in the technology can be scaled. Someday soon, elite diagnostic expertise will be affordably broadcast across the globe, making it available even in regions where people barely have access to any doctor or nurse—let alone to one of the world’s best medical specialists.

现在想象一下，将一项极其具体的创新（基于人工智能的诊断工具或 DeepMind 在蛋白质折叠方面的突破）与几乎无限多的可能性相乘，应用到医学、科学、工业、交通、能源、政府和人类活动的其他所有领域。最终，您将获得一种全新且功能强大的实用工具。本质上是一种“智能电力”。一种灵活的资源，也许有一天只需轻轻一按开关，就可以将认知能力应用于我们面临的几乎所有问题。最终，这种新实用工具不仅能提供分析和决策的能力，还能提供解决复杂问题甚至展现创造力的能力。

Now imagine taking a single, extremely specific innovation—an AI-based diagnostic tool or perhaps DeepMind’s breakthrough in protein folding—and multiplying it by a virtually limitless number of possibilities in other areas from medicine to science, industry, transportation, energy, government and every other sphere of human activity. What you end up with is a new, and uniquely powerful, utility. In essence, an “electricity of intelligence.” A flexible resource that can—perhaps someday with almost a flick of a switch—apply cognitive capability to virtually any problem we face. Ultimately, this new utility will deliver the ability not just to analyze and make decisions but to solve complex problems and even exhibit creativity.

本书的目的是探讨人工智能的未来影响，而不是将其视为一项具体的创新，而是一项独特的可扩展且具有颠覆性潜力的技术——一种强大的新公用事业，有望带来一场有朝一日将与电力相媲美的变革。我将在这里提出的论点和解释主要借鉴了我自己的三次职业经历。

The purpose of this book will be to explore the future implications of artificial intelligence by viewing it not as a specific innovation, but rather as a uniquely scalable and potentially disruptive technology—a powerful new utility poised to deliver a transformation that will someday rival the impact of electricity. The arguments and explanations I will put forth here draw heavily on three of my own professional experiences.

首先，自从我的书《机器人崛起：技术与失业未来的威胁》于 2015 年出版以来，我已受邀在数十场技术会议、地区峰会以及企业和学术活动中谈论人工智能和机器人技术的影响。我去过三十多个国家，有机会参观研究实验室，观看尖端技术的演示，并与技术专家、经济学家、企业高管、投资者和政治家以及看到并开始担心周围正在发生的变化的普通人讨论和辩论正在展开的人工智能革命的影响。

First, since the publication of my book Rise of the Robots: Technology and the Threat of a Jobless Future in 2015, I have been invited to speak about the impact of artificial intelligence and robotics at dozens of technology conferences, regional summits and corporate and academic events. I’ve traveled to more than thirty countries and have had an opportunity to visit research labs, to see demonstrations of leading-edge technology and to discuss and debate the implications of the unfolding AI revolution with technical experts, economists, business executives, investors and politicians, as well as average people who are seeing—and beginning to worry about—the changes happening around them.

其次，2017 年，我开始与法国兴业银行的一个团队合作，创建专有股票市场指数，为投资者提供直接从人工智能和机器人革命中获益的方式。作为咨询主题专家，我帮助制定了一项战略，该战略基于这样的观点：人工智能正在成为一种强大的新工具，因此它将创造价值并改变各行各业的业务。结果是法国兴业银行的“机器人崛起”指数，以及随后基于该指数的Lyxor Robotics 和 AI ETF ^{4 （交易所交易基金）。}

Second, in 2017 I began working with a team at the French bank Société Générale to create a proprietary stock market index that would offer investors a way to benefit directly from the artificial intelligence and robotics revolution. In my role as the consulting thematic expert, I helped formulate a strategy informed by the view that AI is becoming a powerful new utility and that it will therefore generate value and transform businesses in a wide range of industries. The result was Société Générale’s “Rise of the Robots” index and subsequently the Lyxor Robotics and AI ETF⁴ (exchange traded fund), which is based on the index.

最后，在整个 2018 年，我有幸与 23 位全球顶尖的人工智能研究科学家和企业家坐下来进行了广泛的讨论。这些人确实是这个领域的“爱因斯坦”，事实上，与我交谈的人中有四人获得了图灵奖，这是计算机科学领域的诺贝尔奖。这些对话深入探讨了人工智能的未来以及进步将带来的风险和机遇，记录在我 2018 年出版的《智能建筑师：来自人工智能建造者的真相》一书中。我充分利用这个独特的机会，深入了解人工智能领域一些最聪明的头脑，他们的见解和预测直接为本书的大部分内容提供了参考。

Finally, throughout 2018, I had an opportunity to sit down and have wide-ranging discussions with twenty-three of the world’s foremost artificial intelligence research scientists and entrepreneurs. These men and women are truly the “Einsteins” of the field, and indeed, four of the people I spoke with have won the Turing award, computer science’s equivalent of the Nobel prize. These conversations, which delved into the future of artificial intelligence as well as the risks and opportunities that progress will bring, are recorded in my 2018 book Architects of Intelligence: The Truth about AI from the People Building It. I have drawn extensively from this unique opportunity to get inside some of the absolute brightest minds working in the field of AI, and their insights and predictions directly inform much of the material in this book.

将人工智能视为新型电力，为我们提供了一个有用的模型，让我们思考这项技术将如何发展，最终将影响到经济、社会和文化的几乎每一个领域。然而，有一个重要的警告。电力被普遍视为一种毫无疑问的积极力量。除了最虔诚的隐士之外，在发达国家，很难找到任何有理由后悔电气化的人。人工智能则不同：它有阴暗的一面，它给个人和整个社会都带来了真正的风险。

Viewing artificial intelligence as the new electricity offers a useful model for thinking about how the technology will evolve and ultimately touch nearly every sphere of the economy, society and culture. However, there is one important caveat. Electricity is universally viewed as an unambiguously positive force. Setting aside the most dedicated hermit, it would probably be hard to find anyone living in a developed country who has reason to regret electrification. AI is different: it has a dark side, and it comes coupled with genuine risks both to individuals and to society as a whole.

随着人工智能的不断发展，它有可能以前所未有的程度颠覆就业市场和整体经济。几乎任何本质上是常规和可预测的工作——换句话说，几乎任何工人一遍又一遍面临类似挑战的角色——都有可能全部或部分实现自动化。研究发现，多达一半的美国劳动力从事这种可预测的活动，仅在美国，最终就有数千万个工作岗位可能消失。5^而且影响并不仅限于低工资的非技术工人。许多白领和专业人士同样从事相对常规的任务。可预测的脑力劳动尤其容易被自动化取代，因为它们可以由软件完成。相比之下，体力劳动则需要昂贵的机器人。

As artificial intelligence continues to advance, it has the potential to upend both the job market and the overall economy to a degree that is likely unprecedented. Virtually any job that is fundamentally routine and predictable in nature—or in other words, nearly any role where a worker faces similar challenges again and again—has the potential to be automated in full or in part. Studies have found that as much as half of the American workforce is engaged in such predictable activities, and that tens of millions of jobs could eventually evaporate in the United States alone.⁵ And the impact will not be limited to lower-wage, unskilled workers. Many people in white collar and professional roles likewise perform relatively routine tasks. Predictable intellectual work is at especially high risk of automation because it can be performed by software. Manual labor, in contrast, requires an expensive robot.

关于自动化对未来劳动力的影响，人们仍在激烈争论。是否会创造出足够多的新的非自动化工作岗位来吸收那些失去更多常规工作的工人？如果是这样，这些工人是否具备成功过渡到这些新岗位所需的技能、能力和性格特征？我们可能不应该假设大多数前卡车司机或快餐工人可以成为机器人工程师，或者成为老年人的个人护理助理。正如我在《机器人的崛起》中所说，我个人的观点是，随着人工智能和机器人技术的不断发展，我们很大一部分劳动力最终面临被抛在后面的风险。而且，正如我们将看到的，有充分的理由相信，冠状病毒大流行和相关的经济衰退将加速人工智能对就业市场的影响。

There continues to be a vibrant debate over the impact of automation on the future workforce. Will sufficient new, non-automatable jobs be created to absorb the workers who lose more routine work? And, if so, will these workers have the necessary skills, capabilities and personality traits to successfully transition into these newly created roles? We probably should not assume that most former truck drivers or fast food workers can become robotics engineers—or, for that matter, personal care assistants for the elderly. My own view, as I argued in Rise of the Robots, is that a large fraction of our workforce is eventually at risk of being left behind as AI and robotics continue to advance. And, as we’ll see, there are very good reasons to believe that the coronavirus pandemic and the associated economic downturn will accelerate the impact of artificial intelligence on the job market.

即使我们不考虑自动化带来的工作机会的完全消失，技术已经以其他方式影响了就业市场，这应该引起我们的关注。中产阶级的工作面临技术含量降低的风险，因此，一个几乎没有受过培训但通过技术增强的低薪工人可以从事曾经可以要求更高工资的工作。人们越来越多地在算法的控制下工作，这些算法监控或调整他们的工作节奏，实际上是把他们当作虚拟机器人来对待。许多新的机会都来自“零工”经济，在这种经济中，工人的工作时间和收入通常都无法预测。所有这些都表明，我们越来越多的劳动力面临着日益加剧的不平等和可能不人道的条件。

Even if we set aside the complete elimination of jobs through automation, technology is already affecting the job market in other ways that should concern us. Middle class jobs are at risk of being deskilled, so that a low-wage worker with little training, but who is augmented by technology, can step into a role that once would have commanded a higher wage. People are increasingly working under the control of algorithms that monitor or pace their work, in effect treating them like virtual robots. Many of the new opportunities being created are in the “gig” economy, where workers typically have unpredictable hours and incomes. All of this points to increasing inequality and potentially dehumanizing conditions for a growing fraction of our workforce.

除了对就业和经济产生影响外，人工智能的持续崛起还会带来各种其他危险。最直接的威胁之一就是对我们的整体安全。这包括人工智能对物理基础设施和关键系统发起的网络攻击，这些系统将越来越多地通过算法进行互联和管理，以及对民主进程和社会结构的威胁。俄罗斯对 2016 年总统大选的干预为未来可能发生的事情提供了一个相对温和的预兆。人工智能最终可能会通过创建与现实几乎无法区分的照片、音频和视频伪造来增强“假新闻”，而真正先进的机器人大军有朝一日可能会入侵社交媒体，制造混乱，并可能以令人恐惧的熟练程度塑造舆论。

Aside from the impact on jobs and the economy, there are a variety of other dangers that will accompany the continuing rise of artificial intelligence. One of the most immediate threats will be to our overall security. This includes AI-enabled cyberattacks on physical infrastructure and critical systems that increasingly will be interconnected and managed by algorithms, as well as threats to the democratic process and the social fabric. The Russian intervention in the 2016 presidential election offers a relatively tame preview of what might be coming. Artificial intelligence could eventually put “fake news” on steroids by enabling the creation of photographic, audio and video fabrications that are virtually indistinguishable from reality, while armies of truly advanced bots could someday invade social media, sow confusion and potentially mold public opinion with terrifying proficiency.

在世界各地，尤其是在中国，使用人脸识别和其他人工智能技术的监控系统大大增强了专制政府的权力和影响力，并削弱了人们对个人隐私的期望。在美国，人脸识别系统在某些情况下已被证明存在种族或性别偏见，用于筛选简历甚至为刑事司法系统内的法官提供建议的算法也是如此。

Throughout the world—but especially in China—surveillance systems employing facial recognition and other AI-based technologies are being used in ways that vastly enhance the power and reach of authoritarian governments and erode any expectation of personal privacy. In the United States, facial recognition systems have, in some cases, been shown to be biased on the basis of race or gender, as have algorithms used to screen resumes or even to advise judges acting within the criminal justice system.

或许，最可怕的近期威胁是完全自主武器的发展，这种武器无需人类给予特定授权即可杀人。这种武器可能会被大规模使用，以整个人口为目标，而且极难防御，尤其是如果它们落入恐怖分子手中。人工智能研究界的许多人都热衷于阻止这种发展，联合国正在发起一项禁止此类武器的倡议。

Perhaps the most terrifying near-term threat is the development of fully autonomous weapons with the ability to kill without the necessity of a human giving specific authorization. Such weapons could conceivably be used en masse to target entire populations and would be extremely hard to defend against, especially if they fell into the hands of terrorists. This is a development that many people in the AI research community are passionate about preventing, and there is an initiative underway at the United Nations to ban such weapons.

更远的未来，我们可能会遇到更大的危险。人工智能会对人类构成生存威胁吗？我们有一天会不会制造出一台“超级智能”机器，它的能力远远超出我们的能力，以至于它可能会有意或无意地做出伤害我们的事情？这是一种更具推测性的恐惧，只有当我们有一天成功制造出一台真正智能的机器时才会出现。这仍然是科幻小说的内容。尽管如此，创造真正的、人类水平的人工智能是该领域的圣杯，许多非常聪明的人非常重视这一问题。已故的斯蒂芬·霍金和伊隆·马斯克等知名人士已经对失控的人工智能发出了警告，尤其是马斯克，他宣称人工智能研究正在“召唤恶魔”，“人工智能比核武器更危险”，引发了媒体的狂热^。6

Further in the future, we may encounter an even greater danger. Could artificial intelligence pose an existential threat to humanity? Could we someday build a “superintelligent” machine, something so far beyond us in its capability that it might, either intentionally or inadvertently, act in ways that cause us harm? This is a far more speculative fear that arises only if we someday succeed in building a genuinely intelligent machine. This remains the stuff of science fiction. Nonetheless, the quest to create true, human-level artificial intelligence is the Holy Grail of the field, and a number of very smart people take this concern very seriously. Prominent individuals like the late Stephen Hawking and Elon Musk have issued warnings about the specter of out-of-control AI, with Musk in particular setting off a media frenzy by declaring that artificial intelligence research is “summoning the demon” and that “AI is more dangerous than nuclear weapons.”⁶

考虑到所有这些，人们可能会想知道为什么我们应该选择打开潘多拉魔盒。答案是人类不能放弃人工智能。因为人工智能将扩大我们的智力和创造力，它将推动几乎人类所有领域的创新。我们可以预见新的药物和医疗方法、更高效的清洁能源以及许多其他重要突破。人工智能肯定会摧毁工作岗位，但它也会使经济生产的产品和服务更便宜、更易获得。咨询公司普华永道的一项分析预测，到 2030 年，人工智能将为全球经济增加约 15.7 万亿美元——在我们期待从冠状病毒大流行引发的大规模经济危机中复苏之际，这一点就显得更加关键。7^也许最重要的是，人工智能将发展成为一种不可或缺的工具，对于应对我们面临的最大挑战至关重要，包括气候变化和环境恶化、不可避免的下一次大流行、能源和淡水短缺、贫困和缺乏受教育的机会。

Given all this, one might wonder why we should choose to open Pandora’s box. The answer is that humanity cannot afford to leave artificial intelligence on the table. Because AI will amplify our intellect and creativity, it will drive innovation across virtually every field of human endeavor. We can anticipate new drugs and medical treatments, more efficient clean energy sources and a multitude of other important breakthroughs. AI will certainly destroy jobs, but it will also make the products and services produced by the economy more affordable and available. An analysis from the consulting firm PwC predicts that AI will add about $15.7 trillion to the global economy by the year 2030—and this is all the more critical as we look forward to recovery from the massive economic crisis unleashed by the coronavirus pandemic.⁷ Perhaps most importantly, artificial intelligence will evolve into an indispensable tool that will be crucial in addressing the greatest challenges we face, including climate change and environmental degradation, the inevitable next pandemic, energy and fresh water scarcity, poverty and lack of access to education.

未来的道路必须是充分发挥人工智能的潜力，但要保持清醒的头脑。风险需要得到解决。人工智能的具体应用需要受到监管，在某些情况下，甚至需要被禁止。所有这些都需要现在就开始，因为未来早在我们做好准备之前就已经到来了。

The path forward must be to fully embrace the potential of artificial intelligence, but to do so with open eyes. The risks will need to be addressed. Specific applications of AI will need to be regulated and, in some instances, banned. All this needs to begin happening now because the future is poised to arrive long before we are ready for it.

如果说这本书将为人工智能的未来提供“路线图”，那就太夸张了。没有人知道人工智能的发展速度有多快，具体会以何种方式加以利用，会出现哪些新公司和新行业，或者最大的危险是什么。人工智能的未来可能既不可预测又具有颠覆性。没有路线图。我们必须随机应变。我希望这本书能提供一种为未来做好准备的方法：一本思考正在展开的革命的指南，将炒作和耸人听闻与现实区分开来，并找出个人和整个社会在我们正在创造的未来中茁壮成长的最佳方式。

To claim that this book will offer a “roadmap” to the future of artificial intelligence would be to engage in hyperbole. No one knows how rapidly AI will advance, the specific ways in which it will be leveraged, the new companies and industries that will arise or the dangers that will loom largest. The future of artificial intelligence is likely to be as unpredictable as it is disruptive. There is no roadmap. We will have to think on our feet. My hope is that this book will offer a way to prepare for what is to come: a guide to thinking about the unfolding revolution, separating hype and sensationalism from reality, and identifying the best ways for both individuals and our society as a whole to thrive in the future we are creating.

⁷

第二章

CHAPTER 2 人工智能作为新动力

AI AS THE NEW ELECTRICITY

电，一种曾经被人们视为取悦大众的魔术和实验中唯一娱乐来源的力量，无可争辩地塑造和推动了现代文明。在一个人们经常将保证接入电网视为理所当然的世界里，人们很容易忘记电要占据主导地位究竟需要多长时间和多么艰辛。从 1752 年本杰明·富兰克林著名的风筝实验开始，经过了整整 127 年，托马斯·爱迪生才于 1879 年最终发明了白炽灯。从那时起，事情发展得更快了。同年在英国，《利物浦电力照明法案》为该国第一盏电力路灯奠定了基础，仅仅三年后，纽约市的珍珠街发电厂和伦敦的爱迪生电灯站都开始运营。尽管如此，到 1925 年，美国只有大约一半的家庭可以用上电力。又过了几十年，富兰克林·罗斯福颁布了《乡村电气化法案》，电才发展成为今天我们所知的无处不在的公用事业。

ELECTRICITY, A FORCE THAT WAS ONCE VALUED SOLELY AS A SOURCE of entertainment in crowd-pleasing tricks and experiments, has indisputably shaped and enabled modern civilization. In a world where guaranteed access to the electrical grid is often taken for granted, it is easy to forget just how long and arduous electricity’s climb to dominance actually was. From Benjamin Franklin’s famous kite experiment in 1752, a full 127 years passed before Thomas Edison finally perfected his incandescent light bulb in 1879. From that point, things moved faster. That same year in the United Kingdom, the Liverpool Electric Lighting Act laid the groundwork for the country’s first electric street lighting, and just three years later, both the Pearl Street Power Plant in New York City and the Edison Electric Light Station in London began operating. Still, by 1925, only about half of homes in the United States had access to electric power. It took several more decades and Franklin Roosevelt’s Rural Electrification Act before electricity evolved into the ubiquitous utility that we know today.

对于我们这些生活在发达国家的人来说，几乎没有什么事物不受电力影响，或者说电力使之成为可能。电力可能是通用技术的最佳例子，当然也是最持久的例子：换句话说，电力是一种可以跨越并改变经济和社会各个方面的创新。其他通用技术包括蒸汽动力，它引发了工业革命，但现在只限于核电站等少数应用。内燃机无疑是变革性的，但现在很容易想象未来汽油和柴油发动机几乎完全被电动机取代。在没有一些反乌托邦灾难场景的情况下，几乎无法想象没有电的未来。

For those of us who live in the developed world, there is virtually nothing that is not somehow touched by, or indeed made possible by, access to electric power. Electricity is probably the best—and certainly the most durable—example of a general-purpose technology: in other words, an innovation that scales across and transforms every aspect of the economy and society. Other general-purpose technologies include steam power, which produced the Industrial Revolution, but is now relegated to a few applications like nuclear power plants. The internal combustion engine was certainly transformative, but it’s now quite easy to imagine a future where gas and diesel engines are almost entirely displaced—likely by electric motors. In the absence of some dystopian catastrophe scenario, it’s almost impossible to imagine a future without electricity.

因此，认为人工智能将发展成为一种通用技术，其规模和威力足以与电力相媲美，这种说法非常大胆。尽管如此，我们有充分的理由相信这就是我们所走的道路：人工智能就像电力一样，最终将触及并改变几乎所有事物。

It is, therefore, an extraordinarily bold claim to argue that artificial intelligence will evolve into a general-purpose technology of such scale and power that it can reasonably be compared to electricity. Nonetheless, there are good reasons to believe that this is the path we are on: AI, much like electricity, will eventually touch and transform virtually everything.

人工智能已经影响到经济的每一个领域，包括农业、制造业、医疗保健、金融、零售业和几乎所有其他行业。这项技术甚至开始侵入我们认为最人性化的领域。人工智能聊天机器人已经提供全天候心理健康咨询服务。深度学习技术正在产生新的图形艺术和音乐形式。这些都不应该让我们感到惊讶。毕竟，人类创造的几乎所有有价值的东西都是我们智慧的直接产物——我们学习、创新和展现创造力的能力。随着人工智能放大、增强或取代我们自己的智能，它将不可避免地演变成我们最强大、应用最广泛的技术。事实上，人工智能最终可能会被证明是我们从冠状病毒引发的危机中恢复过来的最有效工具之一。

Artificial intelligence is already impacting every sector of the economy, including agriculture, manufacturing, healthcare, finance, retail and virtually all other industries. The technology is even beginning to invade areas that we consider the most human. Already, AI-enabled chatbots provide round-the-clock access to mental health counseling. New forms of graphic art and music are being generated with deep learning technology. None of this should really surprise us. After all, virtually everything of value that human beings have created is a direct product of our intelligence—of our ability to learn, to innovate, to exhibit creativity. As AI amplifies, augments or replaces our own intelligence, it will inevitably evolve into our most powerful and widely applicable technology. Indeed, artificial intelligence may ultimately prove to be one of the most effective tools we have as we look to recover from the crisis unleashed by the coronavirus.

更重要的是，人工智能占据主导地位的速度将远快于电力。这是因为部署人工智能所需的大部分基础设施——包括计算机、互联网、移动数据服务，尤其是亚马逊、微软和谷歌等公司维护的庞大云计算设施——已经到位。想象一下，如果大多数发电厂和输电线路在爱迪生发明灯泡时已经建成，电气化可能会发生得多么快。人工智能即将重塑我们的世界——而且它可能比我们预期的要快得多。

What’s more, it’s a good bet that artificial intelligence will rise to dominance far faster than was the case with electricity. The reason is that much of the infrastructure required to deploy AI—including computers, the internet, mobile data services and especially the massive cloud computing facilities maintained by companies like Amazon, Microsoft and Google—is already in place. Imagine how rapidly electrification might have occurred if most power plants and transmission lines had already been built at the time Edison invented the light bulb. Artificial intelligence is poised to reshape our world—and it may happen much sooner than we expect it.

“智慧之电”

AN “ELECTRICITY OF INTELLIGENCE”

人工智能与电力的类比十分恰当，因为它传达了这样一种感觉：人工智能将无处不在、人人可用，最终将触及和改变我们文明的几乎每个方面。然而，这两种技术之间存在着关键的差异。电力是一种可替代的商品，无论在何时何地都是静态的。无论你身在何处或哪家电力供应商，你通过电网获得的资源基本相同。同样，今天提供的电力与 1950 年相比变化不大。相比之下，人工智能远没有那么同质化，而且更具动态性。人工智能将提供无数不断变化的功能和应用程序，并且可能因技术供应商的不同而有很大差异。正如我们将在第 5 章中看到的那样，人工智能将不懈地继续进步，获得能力并越来越接近人类水平的智能，甚至有朝一日会超越人类。

The analogy to electricity is apt in that it conveys the sense that artificial intelligence will be ubiquitous and universally accessible and that it will ultimately touch and transform nearly every aspect of our civilization. There are, however, critical differences between the two technologies. Electricity is a fungible commodity that is static over both place and time. Regardless of your location or the company that supplies electric power, the resource you access through the electrical grid is essentially the same. Likewise, the electric power on offer today is little changed from what was available in 1950. Artificial intelligence, in contrast, is far less homogeneous and vastly more dynamic. AI will supply myriad and constantly changing capabilities and applications and may vary dramatically based on who exactly is supplying the technology. And as we will see in Chapter 5, artificial intelligence will relentlessly continue to advance, gaining capability and pushing ever closer to human-level intelligence, and perhaps someday beyond.

虽然电力为其他创新提供了动力，但人工智能直接提供了智能——包括解决问题、做出决策的能力，以及未来推理、创新和构思新想法的能力。电力可以为节省劳动力的机器提供动力，但人工智能本身是一种节省劳动力的技术，随着它在整个经济中的应用，它将对人类劳动力以及企业和组织的结构产生巨大影响。

While electricity provides the power that enables the operation of other innovations, AI directly delivers intelligence—including the ability to solve problems, to make decisions and in all likelihood to someday reason, innovate and conceive new ideas. Electricity might power a labor-saving machine, but AI is itself a labor-saving technology, and as it scales across our economy, it will have enormous implications for the human workforce and the structure of businesses and organizations.

随着人工智能不断发展成为一种普遍的实用工具，它将以与电力为现代文明奠定基础相同的方式塑造未来。正如建筑物和其他基础设施的设计和建造是为了利用现有的电网一样，未来的基础设施将从头开始设计以利用人工智能的力量。而且这个想法将超越物理结构，改变我们经济和社会几乎每个方面的设计。新的企业或组织将从一开始就利用人工智能；人工智能将成为未来每种商业模式的关键组成部分。我们的政治和社会机构也将同样发展，以纳入和依赖这种无处不在的新实用工具。

As artificial intelligence continues to evolve into a universal utility, it will shape the future in much the same way that electricity provided a foundation for modern civilization. Just as buildings and other infrastructure are designed and constructed to take advantage of the existing electrical grid, future infrastructure will be designed from the ground up to leverage the power of AI. And this idea will extend beyond physical structures to transform the design of nearly every aspect of our economy and society. New businesses or organizations will be set up to take advantage of AI from their inception; artificial intelligence will become a critical component of every future business model. Our political and social institutions will likewise evolve to incorporate and rely on this ubiquitous new utility.

所有这些的结果是，人工智能最终将达到电力的水平，但它永远不会具有同样的稳定性或可预测性。它将永远是一种更具活力和颠覆性的力量，有可能颠覆它所触及的几乎所有事物。毕竟，智能是终极资源——它是人类所创造的一切的基础能力。很难想象还有什么发展比将这种资源转变为一种普遍可得且负担得起的公用事业更重要。

The upshot of all this is that AI will ultimately achieve the reach of electricity, but it will never have the same stability or predictability. It will always remain a vastly more dynamic and disruptive force with the potential to upend nearly anything it touches. Intelligence is, after all, the ultimate resource—it is the fundamental capability that underlies everything human beings have ever created. It is difficult to imagine a development more consequential than the transformation of that resource into a universally accessible and affordable utility.

新兴硬件和软件人工智能基础设施

THE EMERGING HARDWARE AND SOFTWARE AI INFRASTRUCTURE

和任何公用事业一样，人工智能需要一套支持性基础设施，即一套管道网络，以便让该技术得到普遍应用。当然，这始于已经到位的庞大计算基础设施，包括数亿台笔记本电脑和台式电脑，以及大型数据中心的服务器，以及功能越来越强大的移动设备的快速扩张。通过引入一系列专门用于优化深度神经网络的硬件和软件，这种分布式计算平台作为人工智能交付载体的有效性正在得到显著提高。

Like any utility, artificial intelligence will require an enabling infrastructure, a network of conduits that allows the technology to be universally delivered. This begins, of course, with the vast computing infrastructure already in place, including hundreds of millions of laptop and desktop computers, as well as servers in massive data centers, and a rapidly expanding universe of ever more capable mobile devices. The effectiveness of this distributed computing platform as a delivery vehicle for AI is being dramatically improved by the introduction of a range of hardware and software specifically designed to optimize deep neural networks.

这一演变始于一项发现：专用图形微处理器主要用于实现快节奏视频游戏，是深度学习应用的强大加速器。图形处理单元（GPU）最初旨在增强几乎即时渲染高分辨率图形所需的计算能力。从 20 世纪 90 年代开始，这些专用计算机芯片在高端视频游戏机（如索尼 PlayStation 和微软 Xbox）中尤为重要。GPU 经过优化，可以快速并行执行大量计算。虽然笔记本电脑的中央处理芯片可能有两个或四个计算“核心”，但当代高端 GPU 可能有数千个专用核心，所有这些核心都可以同时高速处理数字。研究人员发现深度学习应用所需的计算与渲染图形所需的计算大致相似，于是他们开始大规模转向 GPU，GPU 迅速发展成为人工智能的主要硬件平台。

This evolution began with the discovery that special graphics microprocessors, used primarily to make fast-action video games possible, were a powerful accelerant for deep learning applications. Graphics processing units, or GPUs, were originally designed to turbocharge the computations required to almost instantaneously render high-resolution graphics. Beginning in the 1990s, these specialized computer chips were especially important in high-end video game consoles, such as the Sony PlayStation and Microsoft Xbox. GPUs are optimized to rapidly perform a vast number of calculations in parallel. While the central processing chip that powers your laptop computer might have two, or perhaps four, computational “cores,” a contemporary high-end GPU would likely have thousands of specialized cores, all of which can crunch numbers at high speed simultaneously. Once researchers discovered that the calculations required by deep learning applications were broadly similar to those needed to render graphics, they began to turn en masse to GPUs, which rapidly evolved into the primary hardware platform for artificial intelligence.

事实上，这一转变是 2012 年开始的深度学习革命的关键推动因素。同年 9 月，多伦多大学的人工智能研究团队在 ImageNet 大规模视觉识别挑战赛中获胜，让深度学习进入了技术行业的视野。ImageNet 大规模视觉识别挑战赛是一项重要的年度机器视觉赛事。如果不依靠 GPU 芯片来加速他们的深度神经网络，获胜团队的作品是否能取得足够好的成绩来赢得比赛是值得怀疑的。我们将在第 4 章中进一步探讨深度学习的历史。

Indeed, this transition was a key enabler of the deep learning revolution that took hold beginning in 2012. In September of that year, a team of AI researchers from the University of Toronto put deep learning on the technology industry’s radar by prevailing at the ImageNet Large Scale Visual Recognition Challenge, an important annual event focused on machine vision. Without relying on GPU chips to accelerate their deep neural network, it’s doubtful that the winning team’s entry would have performed well enough to win the contest. We’ll delve further into the history of deep learning in Chapter 4.

多伦多大学团队使用的 GPU 由 NVIDIA 制造，该公司成立于 1993 年，业务专注于设计和制造最先进的图形芯片。在 2012 年 ImageNet 竞赛之后，深度学习和 GPU 之间的强大协同作用得到了广泛认可，该公司的发展轨迹发生了巨大变化，使其成为与人工智能崛起相关的最杰出的科技公司之一。深度学习革命的证据直接体现在公司的市值上：2012 年 1 月至 2020 年 1 月期间，NVIDIA 的股价飙升了 1,500% 以上。

The University of Toronto’s team used GPUs manufactured by NVIDIA, a company founded in 1993 whose business focused exclusively on designing and manufacturing state-of-the-art graphics chips. In the wake of the 2012 ImageNet competition and the ensuing widespread recognition of the powerful synergy between deep learning and GPUs, the company’s trajectory shifted dramatically, transforming it into one of the most prominent technology companies associated with the rise of artificial intelligence. Evidence of the deep learning revolution manifested directly in the company’s market value: between January 2012 and January 2020 NVIDIA’s shares soared by more than 1,500 percent.

随着深度学习项目迁移到 GPU，领先科技公司的 AI 研究人员开始开发旨在快速实施深度神经网络的软件工具。谷歌、Facebook 和百度都发布了面向深度学习的开源软件，供其他人免费下载、使用和更新。最著名和使用最广泛的平台是谷歌 2015 年发布的 TensorFlow。TensorFlow 是一个全面的深度学习软件平台，为从事实际应用的研究人员和工程师提供优化的代码来实现深度神经网络，以及一系列工具，使特定应用程序的开发更加高效。TensorFlow 和 PyTorch（Facebook 的竞争开发平台）等软件包使研究人员无需编写和测试软件代码来处理晦涩难懂的细节，而是允许他们在构建系统时采用更高层次的视角。

As deep learning projects migrated to GPUs, AI researchers at the leading tech companies began to develop software tools designed to jump-start the implementation of deep neural networks. Google, Facebook and Baidu all released open-source software that was free for others to download, use and update, geared toward deep learning. The most prominent and widely used platform is Google’s TensorFlow, released in 2015. TensorFlow is a comprehensive software platform for deep learning, providing both researchers and engineers working on practical applications with optimized code to implement deep neural networks, as well as a range of tools to make the development of specific applications more efficient. Packages like TensorFlow and PyTorch, a competing development platform from Facebook, free researchers from writing and testing software code to deal with arcane details and allow them to instead take a higher-level perspective as they build systems.

随着深度学习革命的进展，NVIDIA 和一些竞争公司开始开发更强大的微处理器芯片，这些芯片专门针对深度学习进行了优化。英特尔、IBM、苹果和特斯拉现在都设计了带有电路的计算机芯片，旨在加速深度神经网络所需的计算。深度学习芯片正在进入无数应用领域，包括智能手机、自动驾驶汽车和机器人以及高端计算机服务器。其结果是，一个从头开始设计用于提供人工智能的设备网络不断扩大。谷歌于 2016 年宣布推出自己的定制芯片，称为张量处理单元 (TPU)。TPU 专门用于优化使用该公司的 TensorFlow 软件平台构建的深度学习应用程序。最初，谷歌在自己的数据中心部署了新芯片，但从 2018 年开始，TPU 被整合到为公司云计算设施提供支持的服务器中，使得使用其云计算服务的客户可以轻松获得最先进的深度学习功能——这一发展将有助于成为广泛传播人工智能能力的最重要渠道。

As the deep learning revolution progressed, NVIDIA and a number of competing companies moved to develop even more powerful microprocessor chips that were specifically optimized for deep learning. Intel, IBM, Apple and Tesla all now design computer chips with circuitry designed to accelerate the computations required by deep neural networks. Deep learning chips are finding their way into a myriad of applications including smartphones, self-driving cars and robots as well as high-end computer servers. The result is an ever-expanding network of devices designed from the ground up to deliver artificial intelligence. Google announced its own custom chip, called a Tensor Processing Unit or TPU, in 2016. TPUs are specifically designed to optimize deep learning applications built with the company’s TensorFlow software platform. Initially, Google deployed the new chips in its own data centers, but beginning in 2018, TPUs were incorporated into the servers that power the company’s cloud computing facilities, making state-of-the-art deep learning capability easily accessible to clients who utilize its cloud computing service—a development that would contribute to the dominance of what has become the single most important conduit for the widespread distribution of artificial intelligence capability.

为争夺快速增长的人工智能市场份额，老牌微处理器芯片制造商和新一批初创公司展开了激烈竞争，这为该行业注入了活力和创新。一些研究人员正在推动芯片设计朝着全新的方向发展。从 GPU 发展而来的专用深度学习芯片经过优化，可以加快实现深度神经网络的软件执行的复杂数学计算。一种新型芯片更接近于模仿大脑，在很大程度上省去了资源密集型的软件层，并在硬件中实现了神经系统。这些新兴的“神经形态”芯片设计直接在硅片上实例化了神经元的硬件版本。IBM 和英特尔都在神经形态计算研究方面投入了大量资金。例如，英特尔的实验性 Loihi 芯片实现了 130,000 个硬件神经元，每个神经元都可以连接到数千个其他神经元。1^消除大规模软件计算需求的最重要优势之一是能效。人类大脑的能力远远超过任何现有的计算机，但消耗的电量仅为 20 瓦左右，远低于普通白炽灯泡。相比之下，在 GPU 上运行的深度学习系统需要大量电力，而且正如我们将在第 5 章中看到的那样，扩展这些系统以消耗更多资源可能是不可持续的。神经形态芯片的设计直接受到大脑神经网络的启发，耗电量要小得多。英特尔声称，在某些应用中，其 Loihi 架构的能效比传统微处理器芯片高出 10,000 倍。一旦 Loihi 等设计进入商业化生产，它们很可能会迅速融入移动设备和其他以能效为首要考虑因素的应用中。一些人工智能专家走得更远，预测神经形态芯片代表了人工智能的未来。例如，研究公司 Gartner 的一项分析预测，到 2025 年，神经形态设计将在很大程度上取代 GPU，成为人工智能的主要硬件平台^。2

The competition between the established makers of microprocessor chips, as well as a new crop of startups, for a share of the rapidly growing artificial intelligence market has injected a vibrant burst of innovation and energy into the industry. Some researchers are pushing chip designs in entirely new directions. The specialized deep learning chips that evolved from GPUs are optimized to speed up the demanding mathematical calculations performed by software that implements deep neural networks. A new class of chip comes much closer to mimicking the brain, largely dispensing with the resource-hungry software layer and implementing neural systems in hardware. These emerging “neuromorphic” chip designs instantiate hardware versions of neurons directly in silicon. IBM and Intel have both made significant investments in research into neuromorphic computing. Intel’s experimental Loihi chips, for example, implement 130,000 hardware neurons, each of which can connect to thousands of others.¹ One of the most important advantages of eliminating the requirement for software computation at massive scale is power efficiency. The human brain, with capability far beyond any existing computer, consumes only about twenty watts—substantially less than an average incandescent light bulb. Deep learning systems running on GPUs, in contrast, require vast amounts of electricity, and as we’ll see in Chapter 5, scaling these systems to consume ever more resources is likely unsustainable. Neuromorphic chips, with designs directly inspired by the brain’s neural network, are far less power hungry. Intel claims that its Loihi architecture is up to 10,000 times more energy efficient than traditional microprocessor chips in some applications. Once designs like Loihi enter commercial production, they are likely to be quickly incorporated into mobile devices and other applications where power efficiency is a top concern. Some AI experts go much further and predict that neuromorphic chips represent the future of artificial intelligence. One analysis from the research firm Gartner, for example, projects that neuromorphic designs will largely displace GPUs as the primary hardware platform for AI by 2025.²

云计算作为人工智能的主要基础设施

CLOUD COMPUTING AS THE PRIMARY INFRASTRUCTURE FOR ARTIFICIAL INTELLIGENCE

当今的云计算行业始于 2006 年，当时亚马逊网络服务 (AWS) 推出。亚马逊的策略是利用其在构建和管理为其在线购物服务提供支持的大型数据中心方面的专业知识，向广泛的客户出售对托管在类似设施中的计算资源的灵活访问。截至 2018 年，亚马逊网络服务在全球九个不同国家/地区运营着一百多个数据中心。3^亚马逊及其竞争对手提供的云服务增长惊人。根据最近的一项研究，从跨国公司到中小型企业，现在有整整 94% 的组织都在使用云计算。4^到2016 年，AWS 增长如此之快，以至于亚马逊每天必须添加到其系统的新计算资源大致相当于该公司在 2005 年底拥有的所有资源^。5

Today’s cloud computing industry got its start in 2006 with the launch of Amazon Web Services, or AWS. Amazon’s strategy was to leverage its expertise in building and managing the massive data centers that powered its online shopping service by selling flexible access to computing resources hosted in similar facilities to a wide range of clients. As of 2018, Amazon Web Services operated more than one hundred data centers located in nine different countries throughout the world.³ The growth of the cloud services provided by Amazon and its competitors has been staggering. According to one recent study, a full ninety-four percent of organizations, ranging from multinational corporations to small- and medium-sized businesses, now utilize cloud computing.⁴ By 2016, AWS was growing so fast that the new computing resources that Amazon had to add to its system every day were roughly equivalent to everything the company had in place at the end of 2005.⁵

在云提供商出现之前，企业和组织需要购买和维护自己的计算机服务器和软件，并聘请一支高薪技术人员团队来不断维护和升级系统。有了云计算，大部分工作都外包给了亚马逊等提供商，这些提供商能够利用规模经济实现无情的效率。托管云计算服务器的设施通常规模庞大，占地数十万平方英尺，建筑成本高达 10 亿美元，托管超过 50,000 台强大的服务器。云计算资源通常以按需服务的形式提供，客户只需使用并支付在特定时间所需的计算能力、存储和软件应用程序的费用。

Before the advent of cloud providers, businesses and organizations needed to purchase and maintain their own computer servers and software and to employ a team of highly paid technologists to continuously maintain and upgrade the systems. With cloud computing, much of this is instead outsourced to providers like Amazon, who are able to achieve a ruthless level of efficiency by taking advantage of economies of scale. Facilities that host cloud computing servers are typically massive, encompassing hundreds of thousands of square feet in structures costing upward of a billion dollars and hosting more than 50,000 powerful servers. Cloud computing resources are often provided as an on-demand service in which clients utilize and pay for only the computing power, storage and software applications required at any given time.

尽管托管云服务器的设施在物理上规模巨大，但它们严重依赖自动化，因此通常只雇用极少的人员。部署用于管理这些结构内发生的几乎所有事情的复杂算法可以实现在直接人工控制下无法实现的精确度。甚至诸如设施消耗的大量电力以及需要提供冷却以抵消数万台服务器产生的大量热量等因素也经常会随时进行优化。事实上，DeepMind 的 AI 研究的首批实际应用之一就是深度学习系统，它可以优化 Google 自己的数据中心的冷却系统。DeepMind 声称，他们的神经网络是根据从分布在 Google 托管设施中的传感器收集的大量数据进行训练的，能够将用于冷却的能源减少多达 40%。6^算法控制已经产生了真正的好处。 2020 年 2 月发表的一项研究发现，“虽然数据中心的计算量在 2010 年至 2018 年间增长了约 550%，但数据中心在同一时期消耗的能源仅增长了 6%。” ⁷当然，所有这些自动化都会对就业产生影响。向云计算的过渡以及随之而来的大量技术专家职位的消失，这些专家曾经管理着由数千个独立组织维护的计算资源，这可能对 1990 年代后期技术就业热潮的消退做出了重大贡献。

Though the facilities that host cloud servers are physically of massive scale, they rely so heavily on automation that they often employ remarkably few people. Sophisticated algorithms deployed to manage nearly everything that happens inside these structures allow for a level of precision that would be impossible under direct human control. Even factors such as the vast amounts of electrical power consumed by the facilities and the need to provide cooling to offset the massive amounts of heat generated by tens of thousands of servers are often optimized from moment to moment. Indeed, one of the first practical applications of DeepMind’s AI research was a deep learning system that could optimize the cooling systems in Google’s own data centers. DeepMind claims their neural network, which was trained on a trove of data collected from sensors distributed throughout Google’s hosting facilities, has been able to cut the energy used for cooling by up to forty percent.⁶ Algorithmic control has produced real benefits. A study published in February 2020 found that “while the amount of computing done in data centers increased by about 550 percent between 2010 and 2018, the amount of energy consumed by data centers only grew by six percent during the same time period.”⁷ All this automation, of course, has an impact on employment. The transition to cloud computing and the resulting evaporation of huge numbers of jobs held by technical experts who once managed computing resources maintained by thousands of individual organizations likely made a significant contribution to the dampening down of the technology jobs boom that occurred in the late 1990s.

云计算业务模式利润丰厚，各大提供商之间的竞争也十分激烈。AWS 是亚马逊业务中利润最高的部分，利润率远远超过该公司的电子商务业务。2019 年，AWS 的收入增长了 37%，达到 82 亿美元，云服务约占公司总收入的 13%。8 亚马逊的 AWS^仍然是主导力量，约占整个云计算市场的三分之一。微软 2008 年推出的 Azure 服务以及 2010 年推出的 Google Cloud Platform 也占有相当大的市场份额。IBM、中国电子商务巨头阿里巴巴和甲骨文同样是重要的参与者。

The cloud computing business model is highly lucrative and competition among the major providers is intense. AWS is far and away the most profitable part of Amazon’s operations, with margins far in excess of the company’s e-commerce activities. In 2019, revenue from AWS grew thirty-seven percent to $8.2 billion, and the cloud service accounted for about thirteen percent of the company’s total earnings.⁸ Amazon’s AWS remains the dominant force, with roughly a third of the overall cloud computing market. Microsoft’s Azure service, established in 2008, and Google Cloud Platform, launched in 2010, also have significant shares of the market. IBM, the Chinese e-commerce giant Alibaba and Oracle are likewise important players.

政府和企业现在都高度依赖云计算。2019 年，当五角大楼的 JEDI 项目变成一场政治闹剧时，这种依赖所固有的复杂性和党派紧张关系被推到了风口浪尖。JEDI 是联合企业防御基础设施项目的缩写，是一项为期 10 年、价值 100 亿美元的合同，用于托管大量数据并为美国国防部提供软件和人工智能能力。第一次骚动发生在谷歌，当时其员工——他们往往持有政治观点中相当左倾的立场——反对该公司竞标国防相关合同的计划。员工抗议最终导致谷歌退出竞标，该公司在 JEDI 合同竞标截止日期前三天退出了竞标^。9

Governments as well as businesses are now highly dependent on cloud computing. In 2019, the complexities and partisan tensions inherent in this reliance were thrust into the limelight when the Pentagon’s JEDI project turned into a political football. JEDI, an acronym for the Joint Enterprise Defense Infrastructure project, is a ten-year, $10 billion contract to host massive quantities of data and to provide software and artificial intelligence capability to the U.S. Department of Defense. The first kerfuffle occurred at Google, when its employees—who tend to have views positioned pretty far left on the political spectrum—objected to the company’s plans to bid for the defense-related contract. Employee protests eventually led Google to take itself out of the running, and the company withdrew just three days before bids on the JEDI contract were due.⁹

最终，五角大楼将该项目授予了微软 Azure，但由于在该领域的领导地位被视为最有可能的赢家，亚马逊立即声称该决定是出于政治动机。亚马逊于 2019 年 12 月提起诉讼，声称由于唐纳德·特朗普总统对亚马逊首席执行官杰夫·贝索斯的公开敌意，该决定存在不恰当的偏见。贝索斯还拥有《华盛顿邮报》，该报一直对特朗普政府持强烈批评态度。2020 年 2 月，一名联邦法官发布了一项禁令，暂时阻止将合同授予微软。10^一个月后，国防部表示将重新考虑其决定^。11

Eventually, the Pentagon awarded the project to Microsoft Azure, but Amazon, which because of its leadership in the sector was seen as the most likely winner, immediately claimed the decision was politically motivated. Amazon filed a lawsuit in December 2019 claiming the decision was improperly biased because of President Donald Trump’s overt animosity toward Amazon CEO Jeff Bezos. Bezos also owns the Washington Post, which has been highly critical of the Trump Administration. In February 2020, a federal judge issued an injunction temporarily blocking award of the contract to Microsoft.¹⁰ A month later, the Department of Defense said it would reconsider its decision.¹¹

所有这些都生动地说明了云计算市场的竞争必将多么激烈，在某些情况下甚至充满政治风险。而这种竞争态势的核心是人工智能能力，它已成为领先云计算提供商提供的服务中越来越重要的组成部分。深度学习的商业重要性最初是通过科技巨头们努力提供自己的尖端消费者和商业服务来证明的。例如，在内部数据中心的专用硬件上运行的神经网络为亚马逊的 Alexa、苹果的 Siri 以及谷歌的 Assistant 和 Translate 服务提供支持。从这个起点开始，深度学习能力现已完全迁移到这些公司提供的云服务中，并已成为提供商脱颖而出的最重要参数之一。例如，谷歌利用其 TensorFlow 平台的受欢迎程度，为其云客户提供使用其 TPU 芯片构建的强大硬件的直接访问权。反过来，亚马逊利用最新的 GPU 提供深度学习能力，并让其客户运行使用 TensorFlow 或各种其他机器学习平台创建的应用程序。事实上，亚马逊声称，使用谷歌 TensorFlow 开发的云端 AI 应用程序中 85% 实际上是在其自己的 AWS 服务上运行的^。12

All this offers a pretty vivid illustration of just how ferocious, and in some cases politically fraught, the battle for the cloud computing market is certain to be going forward. And at the very center of that competitive dynamic stands the artificial intelligence capability that has become an ever more critical component of the services offered by the leading cloud computing providers. The commercial importance of deep learning was initially demonstrated through the tech giants’ efforts to deliver their own leading-edge consumer and business services. Neural networks running on specialized hardware within internal data centers, for example, power Amazon’s Alexa, Apple’s Siri and Google’s Assistant and Translate services. From this starting point, deep learning capability has now fully migrated into the cloud services offered by these companies, and it has emerged as one of the most important parameters along which the providers differentiate themselves. Google, for example, has leveraged the popularity of its TensorFlow platform by offering its cloud clients direct access to powerful hardware built from its TPU chips. Amazon, in turn, provides deep learning capability utilizing the latest GPUs and lets its clients run applications created using TensorFlow or a variety of other machine learning platforms. Indeed, Amazon claims that eighty-five percent of cloud AI applications developed with Google’s TensorFlow actually run on its own AWS service.¹²

各大云计算公司都在不懈努力，提供更大的灵活性和更好的工具，并迅速应对竞争对手获得的任何优势。英特尔在 2020 年 3 月通过云端提供了一个实验性的神经形态计算系统，这是技术前沿创新的最新例子。该系统由 768 个英特尔类似大脑的 Loihi 芯片构建而成，包含一亿个硬件神经元，大致相当于一只小型哺乳动物的大脑。13如果这种架构被证明是有效的，^那么各大云计算提供商之间的神经形态之战肯定会在短时间内展开。随着各大公司力争超越对方，并在不断增长的面向人工智能的计算资源市场中占据更大份额，其结果是出现了一个从头开始构建的用于提供人工智能的云生态圈。

Among the major cloud companies, there is a relentless drive to offer more flexibility and better tools and to rapidly respond to any advantage gained by a competitor. In one recent example of innovation at the technical frontier, Intel made an experimental neuromorphic computing system available via the cloud in March 2020. The system, built from 768 of Intel’s brain-like Loihi chips, contains one hundred million hardware neurons—roughly equivalent to the brain of a small mammal.¹³ If such architectures prove effective, a neuromorphic battle between the major cloud providers is certain to unfold in short order. As the companies strive to one-up each other and capture a larger share of the ever-increasing market for AI-oriented computing resources, the result has been the emergence of a cloud ecosphere built from the ground up to deliver artificial intelligence.

微软 2019 年向人工智能研究公司 OpenAI 投资了 10 亿美元，OpenAI 与谷歌的 DeepMind 一起引领着深度学习的前沿发展，这为云计算和人工智能之间的自然协同作用提供了一个案例研究。OpenAI 将能够利用微软 Azure 服务托管的大量计算资源，鉴于其专注于构建更大的神经网络，这一点至关重要。只有云计算才能提供 OpenAI 研究所需的规模的计算能力。微软将获得 OpenAI 持续追求通用人工智能所产生的实用创新。这可能会产生可以集成到 Azure 云服务中的应用程序和功能。也许同样重要的是，Azure 品牌将受益于与世界领先的人工智能研究组织之一的合作，并使微软更好地与谷歌竞争，谷歌在人工智能领域享有盛誉，部分原因是它拥有^{DeepMind。14}

Microsoft’s 2019 billion-dollar investment in the AI research company OpenAI—which along with Google’s DeepMind is a leader in pushing the frontiers of deep learning—offers a case study in the natural synergy between cloud computing and artificial intelligence. OpenAI will be able to leverage massive computational resources hosted by Microsoft’s Azure service—something that is essential given its focus on building ever larger neural networks. Only cloud computing can deliver compute power on the scale that OpenAI requires for its research. Microsoft, in turn, will gain access to practical innovations that are spawned by OpenAI’s ongoing quest for artificial general intelligence. This will likely result in applications and capabilities that can be integrated into Azure’s cloud services. Perhaps just as importantly, the Azure brand will benefit from an association with one of the world’s leading AI research organizations and better position Microsoft to compete with Google, which enjoys a strong reputation for AI leadership, in part because of its ownership of DeepMind.¹⁴

这种协同作用远远超出了这一单一示例。从大学研究实验室到人工智能初创公司，再到大型企业正在开发的实用机器学习应用程序，人工智能领域几乎所有重要举措都越来越依赖这种几乎无处不在的资源。云计算可以说是人工智能发展成为一种公用设施的最重要推动因素，而这种公用设施有朝一日将像电力一样无处不在。李飞飞是 ImageNet 数据集和竞赛的架构师，该竞赛成为深度学习革命的催化剂。她从斯坦福大学的现任职位休假，于 2016 年至 2018 年担任谷歌云的首席科学官。她这样说：“如果你想传播人工智能等技术，最好和最大的平台就是云，因为人类发明的任何平台上都没有其他计算能够覆盖如此多的人。谷歌云在任何时候都在赋能、帮助或服务数十亿人。” ¹⁵

This synergy extends far beyond this single example. Virtually every important initiative in AI, ranging from university research labs to AI startups to practical machine learning applications being developed in large corporations, increasingly relies on this nearly universal resource. Cloud computing is arguably the single most important enabler of the evolution of artificial intelligence into a utility that is poised to someday become as ubiquitous as electricity. Fei-Fei Li, the architect of the ImageNet dataset and competition that became a catalyst for the deep learning revolution, took a sabbatical from her current position at Stanford to act as Google Cloud’s chief science officer from 2016 to 2018. She puts it this way: “If you think about disseminating technology like AI, the best and biggest platform is a cloud because there’s no other computing on any platform which humanity has invented that reaches as many people. Google Cloud alone, at any moment, is empowering, helping, or serving billions of people.”¹⁵

工具、培训和人工智能的民主化

TOOLS, TRAINING AND THE DEMOCRATIZATION OF AI

随着新工具的出现，基于云的人工智能向通用工具的演变正在加速，这些新工具使这项技术可供广泛人群使用，而这些人不一定具有很高的技术背景。TensorFlow 和 PyTorch 等平台确实使构建深度学习系统变得更加容易，但它们仍然主要由训练有素的专家使用，这些专家通常拥有计算机科学博士学位。谷歌于 2018 年 1 月推出的 AutoML 等新工具在很大程度上自动化了许多技术细节，并大大降低了进入门槛，让更多人有机会利用深度学习解决实际问题。AutoML 本质上相当于部署人工智能来创造更多人工智能，这是李飞飞所说的“人工智能民主化”趋势的一部分。

The evolution of cloud-based artificial intelligence into a general utility is being accelerated by the emergence of new tools that make the technology accessible to a wide range of people who don’t necessarily have highly technical backgrounds. Platforms such as TensorFlow and PyTorch do make it easier to build deep learning systems, but they are still by and large used by highly trained experts, often with PhDs in computer science. New tools such as Google’s AutoML, introduced in January 2018, largely automate many of the technical details and lower entry barriers substantially, giving far more people the opportunity to utilize deep learning to solve practical problems. AutoML essentially amounts to deploying artificial intelligence to create more artificial intelligence and is part of a trend that Fei-Fei Li calls “the democratization of AI.”

与往常一样，云提供商之间的竞争是创新的强大驱动力，亚马逊为 AWS 平台提供的深度学习工具也变得越来越容易使用。除了开发工具外，所有云服务都提供预构建的深度学习组件，这些组件可以开箱即用并集成到应用程序中。例如，亚马逊提供语音识别和自然语言处理软件包以及一个“推荐引擎”，它可以提出建议，就像向在线购物者或电影观众展示可能感兴趣的替代品一样。16^这种预打包功能最具争议的例子是 AWS 的 Rekognition 服务，它使开发人员可以轻松部署面部识别技术。亚马逊因向执法机构提供 Rekognition 而受到抨击，因为一些测试表明该软件包可能容易受到种族或性别偏见的影响——这是一个道德问题，我们将在第 7 章和第 8章中更^仔细地研究。17

As always, competition between the cloud providers is a powerful driver of innovation, and Amazon’s deep learning tools for the AWS platform are likewise becoming easier to use. Along with the development tools, all the cloud services offer pre-built deep learning components that are ready to be used out of the box and incorporated into applications. Amazon, for example, offers packages for speech recognition and natural language processing and a “recommendation engine” that can make suggestions in the same way that online shoppers or movie watchers are shown alternatives that are likely to be of interest.¹⁶ The most controversial example of this kind of prepackaged capability is AWS’s Rekognition service, which makes it easy for developers to deploy facial recognition technology. Amazon has come under fire for making Rekognition available to law enforcement agencies, given that some tests have suggested the package can be susceptible to racial or gender bias—an ethical issue we will examine more closely in Chapters 7 and 8.¹⁷

第二个重要趋势是在线培训平台的出现，任何具有足够主动性和数学能力的人都可以获得深度学习的基本能力。例如，通过在线教育平台Coursera 提供的 deeplearning.ai 和 fast.ai，后者提供完全免费的在线课程和软件工具，使深度学习更容易获得。18^在就业环境中，进入中上阶层的道路几乎总是需要通过大量时间和金钱的投入来获得正式证书，成为深度学习从业者——至少在当前环境下，对工人的需求远远超过供应——是一个罕见的例外。任何能够成功完成在线课程并展示出使用深度神经网络熟练程度的人，都有机会开启一份收入丰厚、回报丰厚的职业生涯。

A second crucial trend is the arrival of online training platforms that allow anyone with sufficient initiative and mathematical ability to achieve basic competence in deep learning. Examples include deeplearning.ai, which is offered through the online education platform Coursera, and fast.ai, which offers completely free online courses and software tools that make deep learning more accessible.¹⁸ In an employment landscape where the path to the upper middle class nearly always requires formal credentials obtained through massive investments of time and money, becoming a deep learning practitioner—at least in the current environment, where demand for workers far outstrips supply—is a rare exception. Anyone who can successfully complete the online course work and demonstrate proficiency working with deep neural networks has a good shot at launching a lucrative and rewarding career.

随着培训和工具的不断改进，以及越来越多的开发人员和企业家开始部署人工智能应用程序，我们很可能会看到一种寒武纪大爆发，因为这项技术将以无数不同的方式应用。其他主要计算平台也发生了类似的事情。20 世纪 90 年代，当微软 Windows 成为个人电脑的主导平台时，我在硅谷经营一家小型软件公司。最初，Windows 应用程序开发是一项技术性很强的工作，涉及 C 编程语言和数千页的手册，其中充斥着晦涩难懂的细节。易于使用的工具的出现，包括微软的 Visual Basic 等高度可访问的开发环境，极大地增加了可以参与 Windows 编程的人数，并很快导致应用程序的爆炸式增长。移动计算也遵循了类似的轨迹，苹果的 App Store 和 Android Play Store 现在都提供看似无限数量的应用程序来满足几乎所有可以想象到的需求。人工智能，更具体地说是深度学习，很可能也会出现同样的爆炸式增长。在可预见的未来，人工智能作为新电力的出现将由不断扩展的特定应用范围而不是任何更通用的机器智能所推动。

As both training and tools get better and as more developers and entrepreneurs begin to deploy AI applications, we are likely to see a kind of Cambrian explosion as the technology is applied in a myriad of different ways. Something similar has occurred on other major computing platforms. I was running a small software company in Silicon Valley in the 1990s when Microsoft Windows emerged as the dominant platform for personal computers. Initially, Windows application development was a highly technical affair involving the C programming language and thousand-page manuals packed with arcane details. The emergence of easier-to-use tools, including highly accessible development environments like Microsoft’s Visual Basic, dramatically expanded the number of people who could engage in Windows programming and soon led to an explosion of applications. Mobile computing has followed a similar trajectory, and both Apple’s App Store and the Android Play Store now offer a seemingly infinite number of apps to address nearly any conceivable need. The same sort of explosion is likely coming to artificial intelligence, and more specifically to deep learning. The emergence of AI as the new electricity will, for the foreseeable future, be driven by an ever-expanding spectrum of specific applications rather than any more general machine intelligence.

互联世界和“物联网”

AN INTERCONNECTED WORLD AND THE “INTERNET OF THINGS”

“人工智能成为新电力”拼图的最后一块是大幅改善连通性。其中最重要的驱动因素可能是未来几年推出的第五代无线服务 (5G)。5G 预计将使移动数据速度至少提高十倍，甚至可能提高一百倍，同时大幅增加网络容量，从而基本消除瓶颈。19这将不可避免地带来一个更加互联的世界，通信几乎可以即时发生。我们可以想象^，几乎所有东西，包括设备、电器、车辆、工业机械和我们物理基础设施的许多元素，都将互联互通，并通常由运行在云端的智能算法进行监控和控制。这种未来愿景被称为“物联网”，它将引领我们进入一个新世界，例如，冰箱或厨房其他地方的传感器可以检测到您缺少某种物品，然后将该信息传递给算法，该算法会提醒您，甚至可能自动下必要的在线订单。如果冰箱运行不正常，另一种算法通常能够自动或远程解决问题。即将发生故障的部件将被识别并标记以进行更换。将这种模式推广到整个经济和社会，可能会带来巨大的效率提升，因为机器、系统和基础设施会自动诊断并修复出现的问题。物联网在许多方面就像释放目前以超人效率运行云数据中心的算法来运行更广阔的世界。然而，这一切也会带来一些非常现实的风险，特别是在安全和隐私方面，我们将在第 8 章中重点讨论这些关键问题。

The final piece of the “artificial intelligence as the new electricity” puzzle is vastly improved connectivity. The most important driver of this is likely to be the roll out of fifth-generation wireless service (or 5G) in the coming years. 5G is expected to boost mobile data speeds by at least a factor of ten—and perhaps as much as one hundred—while dramatically increasing network capacity so that bottlenecks are largely eliminated.¹⁹ This will lead inevitably to a far more interconnected world where communication happens almost instantaneously. We can imagine that virtually everything—including devices, appliances, vehicles, industrial machinery and a great many elements of our physical infrastructure—will all be interconnected and often monitored and controlled by smart algorithms running in the cloud. This vision of the future has been dubbed the “Internet of Things” and is poised to usher in a world where, for example, sensors in your refrigerator or elsewhere in your kitchen detect that you’re running low on a particular item and then relay that information to an algorithm that alerts you or perhaps even automatically places the necessary online order. If the refrigerator isn’t running optimally, another algorithm will often be able to accomplish an automated or remote resolution. A part that is about to fail will be identified and flagged for replacement. Scaling this model across our entire economy and society is likely to deliver enormous efficiency gains as machines, systems and infrastructure automatically diagnose, and often fix, problems as they arise. The Internet of Things will, in many ways, be like unleashing the algorithms that currently operate cloud data centers with a superhuman level of efficiency to run the wider world. All this will, however, also bring with it some very real risks, especially in the areas of security and privacy, and we will focus on these critical issues in Chapter 8.

这个日益互联的世界将演变成一个强大的人工智能交付平台。在可预见的未来，最重要的人工智能应用将以云为中心。然而，随着时间的推移，机器智能将逐渐变得更加分散。随着设备、机器和基础设施融入最新的专用人工智能芯片，它们将变得越来越智能。神经形态计算等创新可能会产生巨大影响。所有这些的结果是，一个强大的新实用程序几乎可以在任何地方按需提供机器智能。

This ever more connected world will evolve into a powerful platform for the delivery of artificial intelligence. For the foreseeable future, the most important AI applications will be centered in the cloud. However, over time machine intelligence will gradually become more distributed. Devices, machines and infrastructure will become smarter and smarter as they incorporate the latest specialized AI chips. This is where innovations like neuromorphic computing are likely to have a big impact. The upshot of all of this is a powerful new utility carrying the ability to deliver machine intelligence on demand virtually everywhere.

价值在于数据

THE VALUE IS IN THE DATA

随着各大云服务提供商在价格和技术能力方面展开竞争，访问支持人工智能的硬件和软件的成本似乎肯定会下降。与此同时，通过云提供的人工智能服务将不断升级，因为科技巨头们努力通过整合该领域前沿研究人员的最新创新来获得竞争优势。随着这一切的进展，即使是最先进的人工智能技术也将越来越商品化，除了云计算客户为托管数据支付的费用外，几乎不需要任何成本。事实上，已经有证据证明这一点。谷歌、Facebook 和百度等公司都以开源形式发布了他们的深度学习软件；换句话说，他们免费提供这些软件。DeepMind 和 OpenAI 等组织开展的最先进研究也是如此。这两家公司都在领先的科学期刊上公开发表论文，并向所有人公开其深度学习系统的细节。

As the major cloud providers compete on the basis of both price and the capability of their technology, the cost of accessing the hardware and software that enables artificial intelligence seems certain to fall. At the same time, the AI services available through the cloud will be continuously upgraded as the tech giants strive to gain a competitive advantage by incorporating the latest innovations generated by researchers working at the field’s frontier. As all this progresses, even the most advanced AI technologies will become increasingly commoditized and available at little or no cost beyond what cloud computing clients pay to host their data. Indeed, there is evidence of this already. Companies like Google, Facebook and Baidu have all released their deep learning software in open-source form; in other words, they give it away for free. This is also true of the most advanced research conducted by organizations like DeepMind and OpenAI. Both publish openly in leading scientific journals and make the details of their deep learning systems available to everyone.

然而，有一样东西是没有一家公司会免费赠送的：它的数据。这意味着人工智能技术与其消耗的大量数据之间的强大协同作用将不可避免地偏向一个方向。几乎所有产生的价值都将被拥有数据的人所捕获。这一被广泛认可的现实往往导致人们认为科技巨头将完全主宰与大数据或人工智能相交叉的任何领域。然而，这忽略了一个事实，即数据所有权显然是按行业和经济部门垂直划分的。当然，谷歌、Facebook 和亚马逊等公司确实控制着难以想象的海量数据。然而，它通常仅限于网络搜索、社交媒体互动和在线购物交易等领域。在这些领域，老牌公司可能会继续占据主导地位，但更多完全不同类型的数据存在于整个经济和社会中，受政府、组织和其他行业的企业的控制。

There is one thing, however, that no company gives away for free: its data. This means that the powerful synergy between AI technology and the vast quantities of data it consumes will inevitably be skewed in one direction. Nearly all the value generated will be captured by whoever owns the data. This widely recognized reality often leads to the assumption that the tech giants will completely dominate any sphere that intersects with big data or artificial intelligence. However, this overlooks the fact that data ownership is clearly verticalized by industry and economic sector. Companies like Google, Facebook and Amazon do, of course, control unimaginable troves of data. However, it is generally limited to areas like web search, social media interactions and online shopping transactions. In these arenas, the established companies are likely to remain dominant, but far more data of completely different kinds resides across the economy and society, under the control of governments, organizations and businesses in other industries.

人们常说数据是新的石油。如果我们接受这个比喻，那么可以说科技公司在很多方面扮演着与哈里伯顿类似的角色，提供从资源中获取价值所需的技术和专业知识。当然，科技巨头也控制着自己的大量数据储备，但这一不断扩大的全球数据资源的大部分仍掌握在其他人手中。医疗保险公司、医院网络，当然还有政府管理的国家医疗服务等企业控制着价值巨大的数据。可以肯定的是，他们将采用大型科技公司开发并通过云端交付的最新人工智能技术，但他们将在很大程度上保留从数据中提取的价值。金融交易、旅行预订、在线评论、实体零售店内的客户流动以及车辆和工业机械内置的无数传感器产生的运营数据所产生的大量数据也是如此。在每种情况下，机器智能无处不在的新功能都将应用于分布在整个经济中的实体所拥有的特定类型的数据。

It’s often said data is the new oil. If we embrace this analogy, then it’s fair to say that the tech companies in many ways fulfill a role similar to that of perhaps Halliburton, offering the technology and know-how required to extract value from the resource. The tech giants do, of course, also control huge data reserves of their own, but still the lion’s share of this ever-expanding global data resource lies in the hands of others. Businesses like health insurers, hospital networks and, of course, government-managed national health services control data of immense value. To be sure, they will employ the latest AI technology developed by the big technology companies and delivered through the cloud, but they will largely retain the value extracted from their data. The same will be true of the massive amounts of data generated by financial transactions, travel bookings, online reviews, customer movements within physical retail stores and operational data generated by a myriad of sensors built into vehicles and industrial machinery. In each case, the ubiquitous new utility of machine intelligence will be applied to specific types of data owned by entities distributed across the economy.

其中一个重要含义是，人工智能应用所产生的大部分价值将被技术领域中显而易见的候选者以外的实体所获取。利用人工智能带来的巨大利益将广泛传播。再次，用电力来类比是有用的。谁从电力中创造了最大的价值？是电力公司吗？还是核电行业？不，是像谷歌和 Facebook 这样的公司，它们消耗大量电力，并找到了将这种无处不在的商品转化为巨大价值的方法。当然，这个比喻并不完美，毫无疑问，巨大的价值和力量将属于那些在人工智能前沿创新并提供这种不断改进的资源的公司。但是，人工智能应用带来的大部分好处——尤其是当它越来越像商品化的公用事业时——可能会在其他地方产生。

One important implication of this is that much of the value derived from the application of artificial intelligence will be captured by entities beyond the obvious candidates within the technology sector. The enormous benefits from leveraging AI will be distributed widely. Again, an analogy to electricity is useful here. Who generates the most value from electricity? Is it electric utilities? Or the nuclear power industry? No, it’s companies like Google and Facebook that consume massive amounts of electricity and have discovered ways to transform this ubiquitous commodity into fantastic value. The analogy is, of course, not perfect, and without doubt, immense value and power will reside in those companies that innovate on the frontier of artificial intelligence and deliver this ever-improving resource. But most of the benefits that arise from the application of AI—especially as it increasingly resembles a commoditized utility—are likely to accrue elsewhere.

虽然人工智能创造的价值将广泛分布于各个经济部门，但特定行业的情况可能恰恰相反。在商业模式中利用人工智能方面处于前沿的公司可能具有巨大的先发优势。这可能会导致赢家通吃的局面，因为拥有特别有效的大数据和人工智能战略的企业将获得显著的竞争优势。由于数据对于有效应用人工智能至关重要，因此迈向人工智能战略的第一步几乎总是成功的数据战略。这意味着，企业和组织必须专注于建立高效的数据收集和管理系统，作为部署人工智能的前奏。在某些情况下，这将需要解决重要的道德问题，例如有关员工和客户的隐私问题。然而，那些未能积极行动的组织可能会被抛在后面。我们正在迅速走向这样的现实：任何放弃人工智能的企业、政府或组织都是在犯下如此严重的错误，其严重程度可与断开电网相提并论。

While the value created by artificial intelligence will be distributed widely across economic sectors, the reverse may well turn out to be true within a given industry. Companies that are on the frontier when it comes to leveraging AI within their business models are likely to have a substantial first-mover advantage. That could well lead to winner-take-all scenarios as businesses with especially effective big data and artificial intelligence strategies gain a significant competitive advantage. Because data is so central to the effective application of AI, the first step toward an AI strategy is nearly always a successful data strategy. This means that it is crucial for businesses and organizations to focus on building efficient data collection and management systems as a prelude to deploying AI. In some cases, this will require addressing important ethical considerations, for example, around issues of privacy concerning employees and customers. However, those organizations that fail to move aggressively are likely to be left behind. We are rapidly moving toward a reality where any business, government or organization that leaves artificial intelligence on the table is engaging in a misstep of such magnitude that it could reasonably be compared to disconnecting from the electrical grid.

随着人工智能逐渐发展成为一种真正普遍的实用工具，进入每个企业、组织和家庭，它将不可避免地改变我们的经济和社会。这个故事将在未来几年乃至几十年内上演，其影响将不尽相同。在某些领域，人工智能可能会在未来几年带来变革，而在其他情况下，这种颠覆将需要更长的时间才能到来。下一章将探讨人工智能作为一项系统技术的一些实际影响，试图将炒作与现实区分开来，并深入探讨这项快速发展的技术与彻底颠覆我们所有人生活的疫情之间的交集。

AS ARTIFICIAL INTELLIGENCE evolves into a truly universal utility, reaching into every business, organization and household, it will inevitably transform both our economy and our society. This is a story that will play out over the course of years and decades, and the impact won’t be uniform. In some areas AI is likely to be transformative in the next few years, while in other cases the disruption will take much longer to arrive. The next chapter looks at some of the practical implications of artificial intelligence as a systemic technology, attempts to separate hype from reality and delves into the intersection between this rapidly advancing technology and the pandemic that has completely upended all our lives.

¹⁹

第三章

CHAPTER 3 超越炒作：现实主义者眼中的人工智能实用性

BEYOND HYPE: A REALIST’S VIEW OF ARTIFICIAL INTELLIGENCE AS A UTILITY

2019 年 4 月 22 日，特斯拉举办了一场名为“自动驾驶日”的活动。此次活动旨在突出该公司在每辆特斯拉汽车中采用的自动驾驶技术，首席执行官埃隆·马斯克和其他高管及工程师在活动上进行了演讲。马斯克在活动中表示：“我非常有信心预测特斯拉明年将推出自动驾驶机器人出租车。”他还表示，到 2020 年底，特斯拉将在公共道路上拥有 100 万辆这样的汽车。1^马斯克所说的“机器人出租车”是指真正的自动驾驶汽车，能够在无人驾驶的情况下运行，并能够接送乘客并将他们送到随机地点。换句话说，这是 Uber 或 Lyft 的真正机器人版本。

ON APRIL 22, 2019, TESLA HELD AN EVENT IT DUBBED “AUTONOMY Day.” Intended to highlight the autonomous driving technology the company builds into every Tesla, the event featured presentations by CEO Elon Musk and other top executives and engineers. At the event, Musk said, “I feel very confident predicting autonomous robotaxis for Tesla next year.” He went on to suggest that Tesla would have a million such cars operating on public roads by the end of 2020.¹ By “robotaxis,” Musk meant genuine self-driving cars, capable of operating with no one inside and able to pick up passengers and deliver them to random locations. In other words, a truly robotic version of Uber or Lyft.

这是一个令人震惊的预测：与我交谈过的几乎所有其他专家的预期相差甚远。几天后，我出现在彭博电视上，说我对马斯克的预测“感到震惊”，我认为它“过于乐观，甚至有点鲁莽”。我之所以这么说，是因为如此激进的预测几乎肯定会给特斯拉带来市场压力，再加上该公司通过软件下载向特斯拉车主提供新功能的能力，如果未经证实的、声称可以提供完全自动驾驶功能的软件突然交到司机手中，可能会非常危险。虽然一家公司让客户测试新视频游戏或社交媒体应用程序的早期版本可能没问题，但对于显然会导致伤亡的软件来说，这不是一个负责任的策略。i ^{2事实上，已经发生过与特斯拉自动驾驶功能有关的致命事故，该功能可以转向、加速或} ^制动汽车以保持在车道内，但仍需要司机的监督。此外，在我看来，即使该公司能够在一年左右的时间内完善这项技术，也需要更长的时间来充分测试汽车并获得监管部门的批准。因此，到 2020 年底，特斯拉自动驾驶出租车达到 100 万辆是不太可能的。在这段时间内，哪怕只有一辆真正的自动驾驶汽车在公共道路上行驶，也是一件令人惊讶的事情。

This was an astonishing prediction: far out of line with the expectations of virtually every other expert I have talked to. A few days later, I appeared on Bloomberg TV and said that I was “astounded by” Musk’s prediction and that I thought it was “extraordinarily optimistic and perhaps even a bit reckless.” I said this because such an aggressive prediction would almost certainly result in market pressure on Tesla to deliver, and this combined with the company’s ability to provide new features to Tesla owners via software download could be very dangerous if unproven software that purports to deliver fully autonomous capability is suddenly put into the hands of drivers. While it may be fine for a company to have its customers test early versions of a new video game or social media application, this is not a responsible strategy for software that could clearly result in injury or death.ⁱ ² Indeed, there have already been fatal accidents involving Tesla’s autopilot feature, which steers, accelerates or brakes the car to stay within its lane but still requires driver supervision. In addition, it seemed clear to me that even in the unlikely event that the company was able to perfect the technology within a year or so, it would take much longer to adequately test the cars and obtain regulatory approval. So, a million operating Tesla robotaxis by the end of 2020 was just not going to happen. Even a single truly autonomous car operating on public roads within that time frame would be astonishing.

自动驾驶日活动的大部分时间都用于讨论特斯拉正在开发的一款定制的新型自动驾驶微处理器芯片。此前，该公司曾使用过 NVIDIA 生产的针对深度神经网络优化的芯片。特斯拉声称其新芯片提供了前所未有的强大功能，但 NVIDIA 的高管很快反驳道，他们最新版本的 AI 芯片与特斯拉正在开发的产品相当，甚至更快^。3

Much of the Autonomy Day event was devoted to a discussion of a custom new self-driving microprocessor chip being developed by Tesla. Previously the company has used chips optimized for deep neural networks manufactured by NVIDIA. Tesla claimed that its new chip offered unprecedented power, but executives at NVIDIA quickly pushed back, pointing out that the latest versions of their AI chips were equivalent to or even faster than the product under development at Tesla.³

尽管如此，随着自动驾驶日的开展，我清楚地认识到特斯拉确实拥有惊人的竞争优势——最终，这一优势可能使其超越竞争对手，成为第一家部署全自动驾驶汽车的公司。这一优势不是一种特殊的计算机芯片，甚至不是一种算法。相反，正如人工智能领域经常出现的情况一样，这一优势在于特斯拉控制的数据。每辆特斯拉都配备了八个摄像头，它们连续运行，捕捉道路和汽车周围环境的图像。车载计算机能够评估这些图像，确定哪些图像可能引起公司的兴趣，然后自动以压缩格式将这些图像上传到特斯拉的网络。全球有超过 40 万辆配备摄像头的汽车在路上行驶，而且这个数字还在迅速增加。换句话说，特斯拉拥有大量现实世界的照片数据，其竞争对手都无法与之匹敌。

Nonetheless, as I watched Autonomy Day unfold, it became clear to me that Tesla does indeed have a striking competitive advantage—something that ultimately could allow it to outpace its competitors and be the first company to deploy fully autonomous self-driving cars. This advantage is not a special computer chip, or even an algorithm. Rather—as is so often the case in the field of artificial intelligence—the advantage lies in the data that Tesla controls. Every Tesla is equipped with eight cameras that operate continuously, capturing images from the road and the environment around the car. Computers onboard the cars are able to evaluate these images, determine which ones are likely of interest to the company and then automatically upload these in a compressed format to Tesla’s network. Over 400,000 of these camera-equipped cars are driving on roads throughout the world, and that number is increasing rapidly. In other words, Tesla has access to a truly massive trove of real-world photographic data that none of its competitors can come close to matching.

特斯拉人工智能总监 Andrej Karpathy 介绍了该公司如何向配备摄像头的汽车“车队”索取特定类型的图像。例如，如果特斯拉的工程师想要训练其自动驾驶系统以应对道路施工的情况，它可以调取数千张现实世界的建筑工地图像，然后使用这些图像在计算机模拟中训练其自动驾驶软件。虽然所有自动驾驶汽车计划都大量使用模拟，但特斯拉整合大量现实世界数据的能力是一个潜在的颠覆性优势。正如人们常说的那样，事实比小说更离奇，没有一位工程师能够设计出一种模拟系统，能够接近复制特斯拉不断扩张的车队上的摄像头捕捉到的详细且往往怪异的现实。

Tesla’s director of AI, Andrej Karpathy, described how the company can request specific kinds of images from its “fleet” of camera-equipped cars. For example, if Tesla’s engineers want to train its autonomous driving system to handle situations where roads are being worked on, it can summon thousands of real-world images of construction sites and then use those images to train its self-driving software in computer simulations. While all self-driving car initiatives make heavy use of simulation, Tesla’s ability to incorporate massive amounts of real-world data is a potentially disruptive advantage. As it is often said, truth is stranger than fiction, and no engineer could design a simulation that would come close to replicating the detailed, and often weird, reality captured by the cameras on Tesla’s ever-expanding fleet.

这个例子说明了有关人工智能持续进展的新闻往往是一种晦涩难懂的炒作和耸人听闻的混合体，这些新闻交织在传达重要信息的叙述中。正如我所说，人工智能注定会成为一种无处不在的实用工具，最终将触及几乎所有事物。然而，进展不会是同质的：一些技术问题比其他问题更难解决。特别是，一些最引人注目和最受炒作的人工智能应用可能会表现得低于我们的预期，而其他通常不太引人注目的领域的显著进展将让我们措手不及。本章将提供一些示例和指南，以深入了解我认为人工智能可能在相对较短的时间内颠覆哪些领域——以及哪些领域可能需要更长的时间。

This example illustrates how news about ongoing progress in artificial intelligence is often a kind of obtuse mixture of hype and sensationalism woven into a narrative that also conveys important information. As I have stated, artificial intelligence is destined to become a ubiquitous utility that will ultimately touch virtually everything. However, progress will not be homogeneous: some technical problems are much more difficult to solve than others. In particular, some of the highest profile and most-hyped applications of artificial intelligence are likely to underperform relative to our expectations, while dramatic progress in other, often less visible, arenas will take us by surprise. This chapter will present some examples and guidelines that offer insight into the areas where I think AI is likely to be disruptive in the relatively near term—and where it is likely to take much longer.

您的家用机器人交付已延迟

DELIVERY OF YOUR HOME ROBOT HAS BEEN DELAYED

自从第一批科幻小说作者开始对未来进行推测以来，家用个人机器人（既能打扫房间、洗衣服，又能随时充当不知疲倦的管家）的前景就一直吸引着我们的集体想象力。这种机器的前景如何？目前，我们先把我们大多数人都熟悉的真正先进的虚构例子放在一边—— 《杰森一家》中的机器人罗西或《星球大战》中的 C-3PO 之类的人形机器人——考虑一个不那么雄心勃勃的东西：一个功能齐全的机器人，它有实用的（尽管有些有限）能力，可以整理房间，执行各种基本的家庭清洁任务，甚至可能按照命令从冰箱里拿啤酒给我们。我们多久才能看到一个价格合理的个人机器人，我们会觉得它非常有用，甚至不可或缺，以至于大量注重价值的消费者愿意为它买单？

The promise of a personal robot for the home—a machine capable of cleaning the house and doing the laundry even as it stands ready to serve as a tireless butler—has captured our collective imagination since some of the first science fiction authors began to speculate about the future. What are the prospects for such a machine? For the moment, set aside the truly advanced fictional examples most of us are familiar with—Rosie the Robot from The Jetsons or a humanoid machine like C-3PO from Star Wars—and consider something far less ambitious: a functional robot that would have a useful, even if somewhat limited, ability to tidy up a room, perform a variety of basic household cleaning tasks, and perhaps even, on command, bring us a beer from the refrigerator. How soon should we expect to see a reasonably affordable personal robot that we might find so useful, perhaps even indispensable, that masses of value-conscious consumers would be willing to pay for it?

不幸的现实是，这样的机器可能还得等到很远的未来。事实上，迄今为止尝试过的个人机器人的问题在于，它们根本做不了太多事情。真正有用的机器的最低要求，包括在家庭等不可预测的环境中运作所需的视觉感知、移动性和灵活性，是机器人技术中最艰巨的挑战之一。到目前为止，那些试图将消费机器人推向市场的公司甚至还没有真正开始克服这些挑战。相反，他们生产的机器功能非常有限，对于大多数人来说，其价值主张是相当值得怀疑的。

The unfortunate reality is that such a machine probably lies quite far in the future. Indeed, the problem with the personal robots that have been attempted so far is that they simply can’t do very much. The minimal requirements for a truly useful machine, including the visual perception, mobility and dexterity needed to function in an unpredictable environment like a home, are among the most daunting challenges in robotics. So far, the companies that have attempted to bring consumer robots to market haven’t really even begun to overcome these challenges. Instead, they’ve produced machines so limited that, for most people, the value proposition is quite questionable.

一个例子就是 Jibo，这台机器人被誉为首款“社交机器人”。Jibo 是由麻省理工学院的 Cynthia Breazeal 构思的，她是世界上研究机器人的顶尖专家之一，能够与人类进行社交和情感交流。Jibo 于 2017 年秋季推出，是一个大约 12 英寸高的塑料桌面机器人。Jibo 没有手臂、腿或轮子，但它能够倾斜和转动头部，在与主人交流时至少能营造出一种与人类交流的幻觉。该机器人能够进行基本的对话，并能做一些以信息检索为中心的实用事情；它可以在互联网上查找信息、获取天气和交通报告、播放音乐等等。换句话说，Jibo 提供的功能与亚马逊的 Alexa 驱动的 Echo 智能扬声器大致相似。当然，Echo 根本无法移动，但在亚马逊庞大的云计算基础设施和规模更大的高薪 AI 开发人员团队的支持下，它的信息检索和自然语言能力可能会更强大——而且肯定会随着时间的推移而变得更强大。Jibo 最大的问题是其约 900 美元的价格。事实证明，尽管该机器人能够模仿人类的头部动作并随着播放的音乐跳舞，这很可爱也很讨喜，但对于大多数消费者来说，这种能力根本不值得多花 800 美元左右。制造 Jibo 的初创公司在耗费了据称 7000 万美元的风险投资后于 2018 年 11 月倒闭^。4

One example that illustrates these challenges is Jibo, a machine marketed as the first “social robot.” Conceived by MIT’s Cynthia Breazeal, one of the world’s top experts on robots with the ability to engage with people on a social and emotional level, and introduced in the fall of 2017, Jibo is a plastic table-top robot about twelve inches tall. Jibo has no arms, legs or wheels, but it does have the ability to tilt and swivel its head and create at least the illusion of a human-like connection as it communicates with its owner. The robot is able to engage in rudimentary conversations and do a number of practical things that center around information retrieval; it can look up things on the internet, get weather and traffic reports, play music and so forth. In other words, Jibo offers a set of capabilities that are broadly similar to Amazon’s Alexa-powered Echo smart speakers. The Echo, of course, can’t move at all, but backed by Amazon’s massive cloud computing infrastructure and far larger team of highly paid AI developers, it’s information retrieval and natural language capabilities are likely stronger—and certain to become more so over time. Jibo’s biggest problem was its price of around $900. It turned out that, though the robot’s ability to mimic human head gestures and dance along with the music it played was cute and endearing, this capability simply wasn’t worth the extra $800 or so to most consumers. The startup company that made Jibo, after burning through a reported $70 million in venture funding, shut down in November 2018.⁴

据报道，亚马逊正在开发自己的家用机器人。这款机器人代号为 Vesta，被描述为一种“带轮子的 Echo”，能够在家中导航并按照指令到达目的地。5^不过，我还没有看到任何报道称亚马逊计划为这款机器人增加手臂，或者它能够物理操纵其环境。在没有这些功能的情况下，人们再次对它的价值主张感到疑惑。鉴于最便宜的 Echo 版本售价为 50 美元或更低，为什么购买昂贵的移动式（但速度可能很慢）Echo 比在家中各处放置便宜的固定式版本更好呢？这些问题困扰着个人机器人行业，而且目前还不清楚亚马逊是否能够在短期内推出成功的商业产品。

Amazon is reportedly working on its own home robot. Codenamed Vesta, the machine has been described as a kind of “Echo on wheels,” with the ability to navigate around your home and arrive on command.⁵ Still, I have seen no reports that Amazon plans to add an arm to the robot or that it will have any ability to physically manipulate its environment. In the absence of such features, one is again left to wonder about the value proposition. Given that the cheapest versions of the Echo are priced at $50 or less, why would it be better to purchase an expensive mobile (but likely quite slow) Echo, rather than simply place cheap stationary versions throughout your home? These are the kinds of questions that haunt the personal robotics industry, and it’s unclear that even Amazon will be able to launch a successful commercial product anytime soon.

要了解真正实用的家用机器人面临的障碍到底有多大，不妨想象一下一项预期任务：从冰箱里取出一瓶啤酒。假设没有楼梯或关着的门等重大障碍，那么到达冰箱可能就是最容易的部分。机器人在已知环境中导航的技术已经到位，例如 Roomba 机器人吸尘器就证明了这一点。

To get some perspective into how high the hurdle really is for a truly functional home robot, think about just one prospective task: the ability to retrieve a beer from the refrigerator. Assuming there are no major obstacles, such as a staircase or a closed door, getting to the refrigerator is likely the easy part. The technology for robots to navigate within known environments is already in place, as demonstrated, for example, by the Roomba robot vacuum cleaners.

然而，机器人到达后，需要打开冰箱门。您可以亲自尝试一下，看看需要多大的力气。但这不仅仅是蛮力的问题。您可以轻松打开冰箱门，因为您的体重可能超过 100 磅。考虑一下这种情况的物理原理。任何能够成功打开冰箱门的机器人都不是塑料玩具。它不是带轮子的 Amazon Echo。一台除了翻倒之外还能做其他事情的机器需要非常重，而要操纵为人设计的环境，它需要相当接近人体比例。这种机器会很贵。即使我们能找到一种廉价的方法来提供必要的配重 — — 例如，也许给塑料机器人装满水 — — 所需的重量仍然意味着需要强大的马达和重型轮子来推动机器人。

Once the robot arrives, however, it needs to open the refrigerator door. Try this yourself and notice how much force is required. But it’s not just a question of brute strength. You can easily open the door because you likely weigh more than a hundred pounds. Consider the physics of the situation. Any robot that can succeed in opening the refrigerator door is not some plastic toy. It is not an Amazon Echo on wheels. A machine that will do anything other than simply tip over needs to be quite heavy, and to manipulate an environment designed for people, it needs to come reasonably close to human proportions. This machine is going to be expensive. Even if we can figure out a cheap way to provide the necessary counterweight—perhaps filling a plastic robot with water, for example—the required weight still implies the need for a powerful motor and heavy-duty wheels to propel the robot.

门打开后，机器人需要找到啤酒。如果啤酒藏在昨晚晚餐剩下的外卖食品容器后面怎么办？如果啤酒罐被捆在塑料六罐装啤酒环里怎么办？机器人能成功取出一罐吗？想想看，根据剩余啤酒罐的数量，取出啤酒的机制可能会完全不同。是一整罐六罐装啤酒，还是只剩下一个罐子还粘在塑料环上？能够完成这件简单事情的机器人需要非常灵巧，而且可能需要两个非常昂贵的机械臂，而不仅仅是一个。

Once the door is open, the robot needs to locate the beer. What if it’s hidden behind the takeout food containers left over from last night’s dinner? What if the cans of beer are bundled in plastic six-pack rings? Could a robot successfully remove a can? Think about how the mechanics of doing so might be completely different depending on how many cans of beer are left. Is it a full six pack or one lonely can still attached to the plastic? A robot that could do this simple thing would need to be extraordinarily dexterous and would likely need to have two very expensive robotic arms—not just one.

当然，很容易想出一些方法来解决这些问题。也许啤酒需要放在冰箱里正确的位置。忘掉六罐啤酒的包装环吧。罐子必须从包装中取出，也许每个罐子都必须装上 RFID 标签，这样机器人就不必仅仅依靠视觉来找到啤酒。也许有一天，啤酒会采用某种未来主义的包装，这种包装是专门为方便机器人取回而设计的。但就目前而言，所有这些要求都会给您带来不便，从而削弱您花大价钱购买这种机器人的热情。

Of course, it’s easy to imagine ways to get around some of these problems. Maybe the beer needs to be placed in exactly the right location within the refrigerator. Forget six-pack rings. The cans must be removed from any packaging, and maybe each must be fitted with an RFID tag so the robot doesn’t have to rely solely on visual perception to find the beer. Perhaps, someday, beer will come in some sort of futuristic packaging that is specifically designed to make it easy for robots to retrieve it. But for now, all these requirements would add to your inconvenience and therefore diminish your enthusiasm for making a major outlay to acquire such a robot.

毫无疑问，任何真正实用的家用机器人都需要大量的资金投入。电动机、机械臂等部件以及为机器人提供视觉感知、空间定位和触觉反馈所需的各种传感器不受摩尔定律驱动的成本缩减的影响，而摩尔定律是半导体行业的特征，使计算能力变得越来越便宜。家用机器人的根本问题是，为了给消费者提供真正的价值，它至少需要接近我们自己的操控能力。事实证明，人类是令人惊讶的高效生物机器人。

And make no mistake, any genuinely functional home robot will require a very substantial financial investment. Components like electric motors, robotic arms and the various kinds of sensors required to give a robot visual perception, spatial orientation and tactile feedback are not subject to the Moore’s Law–driven cost deflation that has characterized the semiconductor industry and made computing power ever more affordable. The essential problem for a home robot is that, in order to offer real value to consumers, it needs to at least approach our own manipulative capabilities. And human beings, it turns out, are astonishingly effective biological robots.

想象一下您面前的桌子上有两个物体。左边是一个实心钢轴承，直径 3 英寸，重约 4 磅。右边是一个鸡蛋。您可以轻松地拿起这两个物体。想象一下，当您抓住每个物体然后开始举起它时，手部肌肉需要施加的力量。想想如果您以某种方式混淆了物体并施加了错误的力，会有什么后果。即使您蒙着眼睛，您也完全可以仅凭触觉反馈安全地拿起两个物体。在机械手中复制这种能力所需的电机和传感器将是昂贵的——即使有实现这一目标的控制软件。

Imagine two objects on a table in front of you. On the left is a solid steel bearing, three inches in diameter and weighing in at about four pounds. To the right is an egg. You can easily pick up either of these objects. Think about the forces that the muscles in your hand need to apply as you grip each object and then begin to lift it. Consider the consequences if you somehow confused the objects and applied the wrong set of forces. Even if you were blindfolded, it’s a good bet that you would succeed in safely picking up both objects on the basis of tactile feedback alone. The motors and sensors required to replicate that ability in a robotic hand would be expensive—even if the controlling software to make it possible were available.

现实情况是，即使经过数十年对机械手及其动画算法的研究，它们的灵活性仍未接近人类水平。罗德尼·布鲁克斯 (Rodney Brooks) 是世界顶尖机器人专家之一，也是 iRobot Corporation 的联合创始人，该公司是 Roomba 以及一些世界上最先进的军用机器人的制造商，他通过提到经常看到的用于捡拾垃圾的长距离塑料夹持工具来说明这一点：

The reality is that, even after decades of work on robotic hands and the algorithms required to animate them, their dexterity is not yet close to human level. Rodney Brooks, one of the world’s foremost roboticists and a co-founder of iRobot Corporation, the makers of the Roomba as well as some of the world’s most advanced military robots, illustrates this by alluding to the long reach plastic gripper tools that you often see being used to pick up trash:

这种非常原始的 [抓手] 可以进行任何机器人无法做到的奇妙操作，但它只是一块非常原始的塑料垃圾……这就是关键：操作是由你进行的。通常，你会看到研究人员设计的新型机械手的视频，视频中是一个人握着机械手并移动它来执行任务。他们可以用这个小塑料抓取玩具完成同样的任务；这是人类在做的。如果真的那么简单，我们可以将这个抓取玩具连接到机器人手臂的末端并让它执行任务——人类可以用手臂末端的这个玩具完成任务，为什么机器人不能？这缺少了一些戏剧性的东西^。6

That really primitive [gripper] can do fantastic manipulation beyond what any robot can currently do, but it’s an amazingly primitive piece of plastic junk.… That’s the clincher: you are doing the manipulation. Often, you’ll see videos of a new robot hand that a researcher has designed, and it’s a person holding the robot hand and moving it around to do a task. They could do the same task with this little plastic grabber toy; it’s the human doing it. If it was that simple, we could attach this grabber toy to the end of a robot arm and have it perform the task—a human can do it with this toy at the end of their arm, why can’t a robot? There’s something dramatic missing.⁶

即使负责整理房间的机器人能够达到必要的灵活性，它仍然面临着识别可能遇到的数千种不同物体的挑战，然后弄清楚如何处理它们。哪些东西应该小心地放回原位，哪些东西是垃圾，应该丢弃？如果让一个无人看管的机器人在你家的哪怕一个房间里自由活动，你愿意容忍多少错误率？

Even if a robot tasked with tidying up a home were to achieve the necessary level of dexterity, it still faces the challenge of recognizing the many thousands of different objects it might encounter, and then figuring out what to do with them. Which things should be carefully returned to their appropriate place and what items are trash that should be discarded? What error rate would you be willing to tolerate in an unsupervised robot turned loose in even a single room of your home?

这并不是说你的家用机器人永远不会出现。在克服这些障碍方面已经取得了重大进展。例如，未来的机器人似乎可以通过与云交互来识别它们遇到的物体。你已经可以通过谷歌的 Lens 服务看到一个相当令人印象深刻的演示，它允许你将手机指向几乎任何东西并自动生成标识，以及描述信息和类似物体的例子。

None of this is to say that your home robot will never arrive. Significant progress toward overcoming many of these hurdles is already being made. For example, it seems likely that future robots will be able to recognize the objects they encounter by interfacing with the cloud. You can already see a pretty impressive demonstration of this with Google’s Lens service, which allows you to point your mobile phone at nearly anything and automatically generate an identification, as well as descriptive information and examples of similar objects.

随着世界联系越来越紧密，物联网的发展势头越来越强劲，机器人使用的传感器将广泛应用于各种应用，随着对这些设备的需求不断增长，由此产生的规模经济将降低成本。随着机器人越来越多地渗透到商业领域，其他部件最终也可能出现同样的情况。

As the world becomes more connected and the Internet of Things gains traction, sensors of the type used in robots will be deployed widely in a variety of applications, and as demand for these devices grows, the resulting economies of scale should drive down costs. The same is likely to eventually happen with other components as robots increasingly penetrate the commercial sector.

同样，研究人员正在成功部署深度学习和其他技术来构建更灵巧的机械手。最引人注目的演示之一来自 OpenAI，2019 年 10 月，该公司宣布已创建了一个由两个集成深度神经网络组成的系统，使机械手能够解决魔方。7^该系统使用高速模拟进行训练，经过大约 10,000 年的强化学习后才成功。即使对人类来说，用一只手解决魔方也绝非易事。尽管该公司声称它已经实现了“接近人类水平的灵巧性”，但事实证明，对于 OpenAI 的系统来说也并不容易：机械手十次尝试中有八次都掉下了魔方。8^尽管如此，这样的举措代表着真正的进步，正如我们将看到的那样，在许多工业和商业环境中，提高机器人的灵活性将在未来几年内开始产生重大影响。然而，除非在高度不可预测的环境中驱动机器人所需的人工智能得到进一步改进，并且必要的组件变得更便宜，否则，在可预见的未来，价格实惠且真正有用的家用机器人可能仍然遥不可及。

Likewise, researchers are successfully deploying deep learning and other techniques to build more dexterous robotic hands. One of the highest-profile demonstrations came from OpenAI, when in October 2019, it announced that it had created a system consisting of two integrated deep neural networks that enabled a robotic hand to solve a Rubik’s Cube.⁷ The system was trained using high-speed simulation and succeeded only after the equivalent of about 10,000 years of reinforcement learning. Solving a Rubik’s Cube with one hand is not at all easy even for humans. Despite the company’s claim that it had achieved something “close to human-level dexterity,” it turned out that it wasn’t easy for OpenAI’s system either: the robotic hand dropped the cube in eight out of ten attempts.⁸ Still, initiatives like this represent real progress, and as we will see, in many industrial and commercial environments, improving robotic dexterity will begin to have a significant impact within the next few years. Until the artificial intelligence needed to animate robots in highly unpredictable environments gets much better and the necessary components become dramatically cheaper, however, an affordable and genuinely useful robot for the home is likely to remain out of reach for the foreseeable future.

仓库和工厂——机器人革命的发源地

WAREHOUSES AND FACTORIES—GROUND ZERO FOR THE ROBOTIC REVOLUTION

如果说技术限制和经济因素决定了多功能、高效率的家用机器人可能还需要很长时间才能问世，那么在许多工业和商业环境中，情况恰恰相反。在工厂或仓库的封闭空间内，可以消除或至少尽量减少外部世界中的不可预测性和混乱。这通常涉及重新组织设施内人员、机器和材料的互动和流动，以充分利用机器人的功能，同时克服其局限性。严格要求将啤酒（以及所有其他物品）准确放置在冰箱内的精确坐标处，以便机器人可靠地检索，这种要求提供的价值主张似乎并不特别引人注目，但在高容量的商业环境中，即使是微小的效率提高也能带来巨大的经济回报，计算结果却大不相同。

If both technical limitations and economics dictate that a versatile and productive home robot will likely be a long time coming, exactly the opposite is true in many industrial and commercial settings. Within the enclosed space of a factory or warehouse, it’s possible to eliminate, or at least minimize, much of the unpredictability and chaos that rules in the outside world. More often than not, this involves reorganizing the interaction and flow of people, machines and materials within the facility to take advantage of the capabilities of robots, while working around their limitations. The value proposition offered by a strict requirement to unfailingly place the beer—as well as every other item—at precise coordinates within your refrigerator in order to enable reliable robotic retrieval may not seem especially compelling, but in high-volume commercial environments where even a tiny increase in efficiency can result in a massive financial return, the calculation is quite different.

亚马逊和其他在线零售商运营的配送中心的内部运作最能说明这一切。在这些通常规模庞大的设施背后，机器人革命已经在顺利进行中，而且无疑将加速发展。不到十年前，这种仓库里几乎总是有数百名工人在高高的货架之间的过道里不停地穿梭，货架上有成千上万种不同的库存物品。工人通常被分为两组：“堆货员”负责将新到的库存存放在货架上的适当位置，“拣货员”前往这些位置取回物品，以满足客户订单。仓库内的活动将是一场持续不断的疯狂争夺——也许类似于一个特别混乱的蚁丘——一个典型的工人在一个班次中可能要跋涉十几英里甚至更多，匆匆忙忙地往返于随机位置，经常爬上梯子才能到达最上面的货架。

There is no better demonstration of all this than the inner workings of the distribution centers operated by Amazon and other online retailers. Behind the walls of these typically massive facilities, the robot revolution is already well underway and is unquestionably poised to accelerate. Less than a decade ago, warehouses of this kind would have nearly always been animated by hundreds of workers continuously roving the aisles between tall shelves containing thousands of different inventory items. The workers would have been generally divided into two groups: “stowers” tasked with taking newly arrived inventory and storing it at the proper positions on shelves, and “pickers” who traveled to these same locations to retrieve items in order to fulfill customer orders. The activity within the warehouse would have been a continuous mad scramble—perhaps resembling an especially disordered anthill—in which a typical worker might well trek a dozen or more miles over the course of a single shift, hurrying back and forth to random locations and often climbing ladders to reach the uppermost shelves.

在亚马逊最现代化的配送中心，这种忙碌的运动几乎变成了镜像。现在，工人们一动不动，而库存货架则在全自动机器人的背上飞速移动，在目的地之间来回穿梭。这场大规模重组始于亚马逊 2012 年斥资 7.75 亿美元收购仓库机器人初创公司 Kiva Systems。这些机器人看起来有点像巨大的橙色冰球，重达 300 多磅，它们在一个围栏区域内漫游，该区域旨在消除与人类工人发生碰撞的风险，并通过贴在地板上的条形码进行导航。在算法控制下，机器人将装满库存的货架运送到工人所在的工位，然后工人负责将物品存放在可用位置或检索特定产品以满足客户订单。

Within Amazon’s most modern distribution centers, this bustling motion has been transformed almost into a mirror image of itself. It is now the workers who remain stationary, while the inventory shelves speed about, hurrying between destinations riding on the backs of fully autonomous robots. This wholesale reorganization began with Amazon’s $775 million acquisition of the warehouse robotics startup Kiva Systems in 2012. The robots, which look somewhat like huge orange hockey pucks and weigh in at over 300 pounds, roam within a fenced-off area designed to eliminate any risk of collisions with human workers and navigate by following barcodes attached to the floor. Operating under algorithmic control, robots deliver shelves laden with inventory to stations occupied by workers, who are then tasked with either stowing items at an available location or retrieving a specific product to fulfill a customer order.

亚马逊目前在全球各地的配送中心运营着 20 多万台这样的机器人。这使得一名普通拣货员在一小时内可以取回的物品数量增加了三到四倍。9^{到目前为止}，机器人在很大程度上并没有取代工人。事实上，亚马逊仓库的就业人数大幅增加——在一定程度上抵消了传统零售环境中随着网上购物的兴起而消失的就业岗位。机器人在光滑、无障碍的地面上快速行走，可携带重达 700 磅的库存，而工人则留在原地，执行需要视觉感知和灵活性的任务，至少到目前为止，这些任务超出了任何机器人的能力。10^工人和机器之间的这种协同作用有助于亚马逊不断提高其为客户提供的服务水平。例如，如果没有对机器人的大规模投资，亚马逊于 2019 年推出的一日送达服务很可能是不可能实现的。同样，自动化可能至关重要，因为随着冠状病毒危机的爆发，尽管亚马逊的许多仓库工人都病倒了，但它仍努力跟上不断飙升的需求。

Amazon now operates more than 200,000 of these robots at its distribution centers worldwide. The result has been a three- to four-fold increase in the number of items that can be retrieved by a typical picker over the course of an hour.⁹ So far, the robots have not, for the most part, replaced workers. Indeed, employment in Amazon warehouses has grown dramatically—offsetting to some extent the evaporation of jobs in traditional retail settings as online shopping has gained traction. The robots rapidly navigate smooth, obstacle-free floors carrying up to 700 pounds of inventory, while workers stay in place, performing tasks requiring the visual perception and dexterity that, so far at least, is beyond the capability of any robot.¹⁰ This synergy between workers and machines has been instrumental in enabling Amazon to continuously ramp up the level of service it offers its customers. One-day delivery for Amazon’s Prime customers, introduced in 2019, for example, would have likely been impossible without this massive investment in robotics. Likewise, automation was likely crucial as Amazon struggled to keep pace with soaring demand as the coronavirus crisis unfolded, even as many of its warehouse workers fell sick.

虽然这种将工人和机器人结合起来的方式充分利用了各自的相对优势，带来了不可否认的效率提升，但它也改变了这些工作的性质，既有积极的一面，也有消极的一面。在新的制度下，在仓库过道中跋涉的疲惫被令人麻木的重复工作所取代。工人们现在站在原地，一小时又一小时地将货物从到达的货架上存放或挑选出来。根据一项分析，亚马逊仓库的受伤率是仓库行业平均水平的两倍多，而且随着新机器人技术的出现，受伤率实际上有所增加，部分原因是重复性动作带来的痛苦或从较高的货架上取回重物的压力。¹¹行业顾问马克·伍尔夫拉特 (Marc Wulfraat) 告诉 Vox 记者杰森·德尔雷 (Jason Del Rey)：“每天在水泥地上行走 12 英里来挑选这些订单……如果你还不到 20 岁，那么到周末你就已经筋疲力尽了……使用橡胶垫来送货上门，比传统方式生产率高出三倍，也更人性化……[但]三倍的挑选速度也意味着更多的磨损，因为需要重复动作，并且需要更快地抬起和搬运产品。” ¹²

While this coupling of workers and robots in ways that leverage the relative strengths of each results in undeniable efficiency gains, it also transforms the nature of these jobs in ways that can be both positive and negative. Under the new regime, the exhausting trudge through the warehouse aisles has been replaced with mind-numbing repetition. Workers now stand in place and stow or pick items from the arriving shelves hour after hour. According to one analysis, warehouse injuries at Amazon occur at more than twice the warehouse industry average and have actually increased with the new robotic technology in part because of repetitive motion distress or the strain of retrieving heavy items from higher shelves.¹¹ As Marc Wulfraat, an industry consultant, told Vox reporter Jason Del Rey, “Walking 12 miles a day on a concrete floor to pick these orders.… If you’re not 20 years old, you’re a broken person at the end of the week.… Having a rubber mat, where goods come to you, is three times more productive than the traditional approach and more humane,… [but] picking three times faster also implies more wear and tear due to repetitive motion and working faster at lifting and handling products.”¹²

事实是，在这样的设施中，工人逐渐丧失了自主权，变成了本质上相当于插入式生物神经网络的东西，通过提供迄今为止机器智能无法实现的功能，填补了高度机械化过程中的空白。结果之一是美国和欧洲配送中心的抗议活动，人们抱怨人类被当成机器人一样对待，工人们在越来越苛刻的算法的监督下不断被迫满足不合理的期望。13在我看来^，如果这些工作越来越被认为是非人性化的，甚至是危险的，随着工人越来越接近身体和心理的极限，一旦必要的技术出现，消除这些工作将不可避免地成为一种理性行为。

The truth is that, within facilities like this, workers are gradually losing their agency and being transformed into what essentially amounts to plug-in biological neural networks that fill the gaps in a largely mechanized process by rendering the capabilities that are—so far—beyond the reach of machine intelligence. One result has been protests at distribution centers in both the United States and Europe, complaining that human beings are being treated like robots and that workers are continuously driven to meet unreasonable expectations under the supervision of ever more demanding algorithms.¹³ It seems to me that if these jobs are increasingly perceived as dehumanizing or even dangerous, and as workers are pushed ever closer to their physical and psychological limits, it will inevitably become a rational for their elimination as soon as the necessary technology arrives.

事实上，在这种封闭且相对受控的环境中，自动化的棘轮可能会无情地向前推进，逐渐推动运营变得越来越不劳动密集。亚马逊已经在积极推动其仓库运营的更多方面实现自动化。路透社记者杰弗里·达斯汀 (Jeffrey Dastin) 在 2019 年 5 月的一篇报道中透露，亚马逊一直在引进先进的机器，这些机器能够将产品最终包装到盒子中，准备发送给客户。由于机器人仍然缺乏可靠的灵活性，无法可靠地拾取高度多样化的产品并将其放入盒子中，因此机器可以在物品沿着传送带移动时几乎立即围绕物品构建定制尺寸的盒子。这些机器每小时可以装箱约 600 到 700 件物品——相当于人类工人能力的五倍。两名参与亚马逊该项目的人士告诉达斯汀，这最终可能导致美国多达 55 个仓库中约 1,300 个工作岗位消失^。14

Indeed, within these kinds of enclosed and relatively controlled environments, the ratchet of automation is likely to click forward relentlessly, gradually driving operations to become ever less labor intensive. Amazon is already moving aggressively to automate more aspects of its warehouse operations. A May 2019 report by Reuters journalist Jeffrey Dastin revealed that Amazon has been introducing advanced machines capable of doing the final packaging of products into boxes ready to ship to customers. Because robots still lack the dexterity to reliably pick up highly varied products and place them in boxes, the machines instead almost instantly construct custom-sized boxes around an item as it travels along a conveyor belt. The machines can box about 600 to 700 items per hour—as much as five times the capability of a human worker. Two people who had been involved in the project at Amazon told Dastin that this could ultimately lead to the elimination of about 1,300 jobs in as many as fifty-five warehouses located throughout the United States.¹⁴

同样，亚马逊在分拣中心引入了类似曲棍球形 Kiva 机器人的小型机器人，该公司在分拣中心将包裹分发给开往不同目的地的卡车。小型机器人不是搬运货架上的库存，而是将单个包裹运送到分拣中心地板上的特定位置（对应于邮政编码），然后包裹滑入地板上的一个洞中，并被送到下面等候的卡车上。15^当然，所有这些都提供了一个生动的例子，说明如何从头开始设计和重组整个工作环境，以最大限度地发挥机器人自动化强大但有限的能力。随着机器人的进步和变得更加灵活和熟练，这些环境肯定会定期进行重组，以利用新的可能性并最大限度地提高生产力。

Likewise, Amazon has introduced robots that look like smaller versions of its hockey puck–shaped Kiva robots in the sorting centers where the company distributes packages to trucks headed to various destinations. Rather than carrying shelves of inventory, the smaller robots instead tote a single package to a specific location on the sorting center floor—corresponding to a zip code—where the package then slides into a hole in the floor and is sent on its way to a truck waiting below.¹⁵ All this, of course, offers another vivid example of how an entire work environment can be designed and restructured from the ground up to maximize the powerful, albeit limited, capabilities of robotic automation. As robots advance and become more versatile and adept, these environments are sure to be periodically restructured to take advantage of new possibilities and maximize productivity.

在仓库和工厂中，一旦机器人在抓取和操纵物体的能力方面最终接近人类水平，自动化的终局就会展开。超过这个水平，一个完全自动化的仓库将成为一个现实场景，其中就业人数仅限于监督和维护机器的相对较少的工人。亚马逊显然对实现这一里程碑表现出浓厚的兴趣；该公司组织了一系列备受吹捧的年度竞赛，来自世界各地大学的工程团队竞相制造机器人，这些机器人可以执行目前由工人在仓库货架上挑选物品的任务。16^虽然制造一只能够可靠地抓取数千种不同物品（所有物品的尺寸、重量、形状、质地和包装配置都不同）的机械手已被证明是一项艰巨的挑战，但沿着这条道路前进是不可避免的。亚马逊首席执行官杰夫·贝佐斯 (Jeff Bezos) 在 2019 年 6 月的一次会议上表示，“我认为抓取将在未来 10 年内得到解决”，尽管“事实证明这是一个极其困难的问题，可能部分原因是我们开始用机器视觉来解决它，所以机器视觉必须放在首位。” ¹⁷换句话说，目前从事库存堆垛工和拣货工的数千人（占公司仓库员工的大多数）可以说在十年左右的时间内就会被裁员。

Within warehouses and factories, the automation endgame will unfold once robots finally do approach human-level capability in terms of their ability to grasp and manipulate objects. Beyond this point, the specter of a fully automated warehouse where employment is limited to a relatively small number of workers who supervise and maintain the machines becomes a realistic scenario. Amazon has clearly demonstrated a keen interest in achieving this milestone; the company has organized a number of highly touted annual contests in which engineering teams from universities around the world have competed to build robots that can perform the tasks now undertaken by the workers who pick items from the shelves in its warehouses.¹⁶ While building a robotic hand capable of reliably grasping thousands of different items—all of different sizes, weights, shapes, textures and packaging configurations—has proven to be a daunting challenge, progress along this path is inevitable. Speaking at a conference in June 2019, Amazon CEO Jeff Bezos said, “I think grasping is going to be a solved problem in the next 10 years,” in spite of the fact that “it’s turned out to be an incredibly difficult problem, probably in part because we’re starting to solve it with machine vision, so machine vision did have to come first.”¹⁷ In other words, the thousands of people currently engaged as inventory stowers and pickers—a majority of the company’s warehouse workforce—are arguably on a pretty clear glide path toward being made redundant within a decade or so.

然而，很有可能，对就业的影响将在那之前很久就开始显现。同样，关键因素是仓库内受控、相对可预测的环境。在我看来，在这样的环境中，机器人不必近乎完美，就能增加显著的价值。事实上，只要机器人始终以可预测的方式出现故障，那么能够可靠地处理典型仓库中 50% 甚至更少的库存物品的机器人就可以大幅提高生产率。亚马逊拥有大量数据流，可用于准确预测履约机器人可能成功和可能失败的地方。从客户下在线订单的那一刻起，公司当然就知道涉及哪些物品，并且应该很容易预测该订单是否适合完全由机器人执行，或者是否需要将其发送给人类工人。换句话说，亚马逊只需管理其配送中心内的工作流程，就可以利用能力有限的机器人。

In all likelihood, however, the impact on jobs will begin to manifest long before then. Once again, the key factor is the controlled, relatively predictable environment within a warehouse. It seems to me that in a setting like this, a robot does not have to be anywhere near perfect in order to add significant value. Indeed, a robot that can reliably handle fifty percent—or perhaps or even less—of the items inventoried in a typical warehouse could drive massive productivity increases, as long as the robot consistently fails in a predictable way. Amazon has at its disposal enormous streams of data that can be used to predict exactly where a fulfillment robot is likely to succeed and where it might fail. From the instant a customer places an online order, the company knows, of course, exactly what items are involved and should have little trouble predicting whether that order is a good candidate for fully robotic fulfillment or if it instead needs to be routed to a human worker. In other words, Amazon can take advantage of robots with limited capability simply by managing the flow of work within its distribution centers.

这种能够准确预测机器人操作结果并避免故障的能力实际上是受控仓库环境与更加混乱的外部世界之间的分界线，在受控仓库环境中，机器人很可能在不久的将来蓬勃发展，而在更加混乱的外部世界，自动驾驶汽车等技术面临的挑战可能更加艰巨。如果仓库机器人能够可预测地处理一半可能遇到的物品，那么它就非常有用。如果自动驾驶汽车在公共道路上行驶，能够可靠地应对 99% 遇到的情况，那么它可能比无用更糟糕，因为那 1% 几乎注定会酿成灾难。

This ability to reliably predict the outcome of a robotic operation and work around failure is really the bright line between a controlled warehouse-type environment, where robots are likely to thrive in the relatively near future, and the far more chaotic outside world, where the challenges for technologies like self-driving cars are likely to be far more daunting. A warehouse robot that can predictably handle half the items it might encounter can be highly useful. A self-driving car operating on a public road that can reliably navigate ninety-nine percent of the situations it encounters may be worse than useless because that outlying one percent virtually guarantees disaster.

亚马逊的销售受长尾分布的支配，仓库中库存的一小部分产品构成了客户订购的大部分商品，这一事实很可能使部分能力的履行机器人更有价值。能够持续抓取和操纵这些受欢迎、大批量商品的很大一部分的机器人将是提高生产率的特别有效的方法。当然，没有机器人是完全可靠的，即使只负责履行那些预计能够处理的订单。为了处理相对罕见的故障，很容易想象一个人类工人监督几个履行机器人的运行，只在出现问题时进行干预。结果是，仓库自动化不是在机器人真正达到人类水平的灵活性后才大规模实现的，而是更有可能逐步、零碎地发展，其中流程的每个阶段都可能需要对仓库内的工作流程进行重大重组。

A partially capable fulfillment robot is likely made even more valuable by the fact that Amazon’s sales are governed by a long-tail distribution in which a relatively small fraction of the products stocked in a warehouse constitute the lion’s share of the items that customers tend to order. A robot with the ability to consistently grasp and manipulate a significant percentage of these popular, high-volume items would be an especially effective way to achieve productivity gains. Of course, no robot will be completely reliable, even when tasked with fulfilling only those orders that it is expected to be able to handle. To deal with relatively rare failures, it’s easy to imagine one human worker overseeing the operation of several fulfillment robots and interceding only when a problem occurs. The upshot is that rather than warehouse automation arriving en masse only after truly human-level robotic dexterity is achieved, it’s more likely to take place gradually, in a piecemeal evolution, in which each stage of the process may require a significant reorganization of the workflow within warehouses.

超越亚马逊，探索真正的灵巧机器人

BEYOND AMAZON AND THE QUEST FOR TRULY DEXTEROUS ROBOTS

虽然亚马逊在机器人领域的举措因其规模和影响力而备受关注，但其在线竞争对手以及各种实体零售连锁店运营的设施的情况大致相同。尤其是北美和欧洲的杂货店正在积极向配送中心自动化迈进，以提高效率并深入开展在线销售，部分原因是预计亚马逊在 2017 年 6 月收购全食超市后将颠覆食品杂货市场。

While Amazon’s initiatives in robotics attract a great deal of attention because of the company’s size and influence, the story is broadly similar at facilities run by its online competitors as well as a variety of brick-and-mortar retail chains. Grocery stores in both North America and Europe, in particular, are aggressively moving toward automation in distribution centers as a way of becoming more efficient and delving into online sales, partly in anticipation of Amazon’s looming disruption of the grocery market in the wake of its acquisition of Whole Foods in June 2017.

英国的 Ocado 是该领域的领军企业之一，该公司经营着自己的在线杂货服务，并向全球超市连锁店推销其仓库自动化技术。在该公司位于英国安多弗的配送中心，一千多个机器人在轨道上运行，轨道排列成一个高架网格结构，类似于一个巨大的棋盘。最多可存放 25 万个板条箱（每个板条箱都装有特定的杂货），存放在与棋盘上的方格相对应的位置。机器人在上面的轨道上导航，抓住板条箱，然后将其拉入其类似盒子的内部，并将它们运送到工位，在那里检索单个物品并打包客户订单。机器人自主运行，通过移动数据网络相互通信和导航，并定期返回对接站为电池充电。18^甚至还有专门的检索机器人，如果其中一个板条箱搬运机器人发生故障，它们会前来救援。安多弗工厂每周可处理约 65,000 份在线杂货订单，包含 350 万件单品^。19

One of the leaders in this arena is United Kingdom–based Ocado, which runs its own online grocery service and also markets its warehouse automation technology to supermarket chains worldwide. At the company’s distribution center in Andover, England, more than a thousand robots run on rails arranged in an elevated grid structure resembling a massive checkerboard. Up to 250,000 crates—each stocking a particular grocery item—can be stored at locations corresponding to squares on the board. The robots navigate on the rails above, grappling and then pulling crates up into their box-like interiors and carrying them to stations where individual items are retrieved and customer orders packed. The robots run autonomously, communicating and navigating around each other via a mobile data network, and periodically returning to docking stations to have their batteries charged.¹⁸ There are even specialized retrieval robots that come to the rescue in the event that one of the crate-carrying robots malfunctions. The Andover facility is capable of processing about 65,000 online grocery orders, containing three and half million individual items, every week.¹⁹

和亚马逊的仓库一样，机器人专注于快速移动物料的物流，而人类在所有这些自动化系统中的主要作用是挑选和包装，这仍然需要人类的灵活性。典型的购物清单上有各种各样的物品——从罐头食品到盒装食品再到新鲜农产品——这对机器人的操作提出了特别的挑战。正如科技记者詹姆斯·文森特 (James Vincent) 指出的那样，“没有什么比一袋橙子更让机器人难倒了。”困难在于“袋子的移动方式太多奇怪，没有明显的部分可以抓住，如果你用力挤压，最终得到的会是橙汁。” ²⁰尽管如此，Ocado 已经在尝试使用机器人来克服这些挑战。该公司正在使用机器人拾取臂，使用吸盘来举起具有合适表面的物品，例如罐头，以及软橡胶机器人手，有朝一日它们将能够有效地抓住更易碎的物品。

As in Amazon’s warehouses, the robots focus on the logistics of rapidly moving materials, while the primary role for humans among all this automation is the picking and packing that continues to require human dexterity. The wide variety of items that characterizes a typical grocery list—everything from canned goods to boxed items to fresh produce—presents a particular challenge for robotic manipulation. As technology journalist James Vincent points out, “nothing stumps a robot quite like a bag of oranges.” The difficulty is that “the bag moves in too many weird ways, there are no obvious bits to grab hold of, and if you squeeze too hard you end up with orange juice instead.”²⁰ Nonetheless, Ocado is already experimenting with robots that attempt to overcome these challenges. The company is employing robotic picker arms that use suction cups to lift items with suitable surfaces, such as cans, as well as soft rubber robotic hands that someday will be able to effectively grasp more fragile items.

打造真正灵巧的机器人已成为硅谷风险投资公司的一大关注点，许多资金雄厚的初创公司纷纷涌现，采用各种方法在研究前沿进行创新。其中最引人注目的初创公司之一是 Covariant，该公司成立于 2017 年，但直到 2020 年初才从隐身模式中脱颖而出。Covariant 的研究人员认为，“强化学习”——本质上是通过反复试验进行学习——是最有效的进步方式，该公司声称正在构建一个基于大规模深度神经网络的系统，称之为“机器人通用人工智能”，预计该系统最终将为各种机器提供动力，这些机器可以“观察、推理和行动周围的世界，完成传统编程机器人无法完成的复杂多样的任务”。²¹该公司由加州大学伯克利分校和 OpenAI 的研究人员创立，获得了深度学习领域一些顶尖人物的投资和宣传，其中包括图灵奖获得者 Geoffrey Hinton 和 Yann LeCun、谷歌的 Jeff Dean 和 ImageNet 创始人李飞飞。22 ²⁰¹⁹年，Covariant 在瑞士工业机器人制造商 ABB 组织的一场竞赛中击败了其他 19 家公司，展示了唯一能够识别和操作各种物品而无需人工干预的系统。23 Covariant 将与 ABB 以及其他大公司合作，为部署在仓库和工厂的工业机器人注入智能，该公司相信这种智能最终将^达到或超过人类的感知和灵活性。

The quest to build a truly dexterous robot has become a major focus of Silicon Valley venture capital firms, and a number of well-funded startup companies have emerged, embracing varied approaches as they innovate at the research frontier. One of the highest-profile startups is Covariant, which was founded in 2017 but emerged from stealth mode only in early 2020. The researchers at Covariant believe that “reinforcement learning”—or essentially learning through trial and error—is the most effective way to progress, and the company claims to be building a system based on a massive deep neural network that it calls “universal AI for robots,” which it expects to eventually power a variety of machines that can “see, reason and act in the world around them, completing tasks too complex and varied for traditional programmed robots.”²¹ The company, founded by researchers from the University of California at Berkeley and OpenAI, has received investments and publicity from some of the brightest lights in deep learning, including Turing Award winners Geoffrey Hinton and Yann LeCun, Google’s Jeff Dean and ImageNet founder Fei-Fei Li.²² In 2019, Covariant defeated nineteen other companies in a competition organized by the Swiss industrial robot maker ABB by demonstrating the only system capable of recognizing and manipulating a variety of items without any need for human intervention.²³ Covariant will be working with ABB as well other major companies to imbue industrial robots deployed in warehouses and factories with intelligence that the company believes will eventually match or exceed human-level perception and dexterity.

许多在该领域工作的初创公司和大学研究人员都和 Covariant 一样认为，以深度神经网络和强化学习为基础的战略是推动更灵巧的机器人进步的最佳方式。但有一个显著的例外，那就是位于旧金山湾区的一家小型人工智能公司 Vicarious。Vicarious 成立于 2010 年（2012 年 ImageNet 竞赛将深度学习推向前台的两年前），其长期目标是实现人类水平的人工智能或通用人工智能。换句话说，从某种意义上说，该公司正在与 DeepMind 和 OpenAI 等知名度更高、资金更充足的项目直接竞争。我们将在第 5 章中深入探讨这两家公司正在开辟的道路以及对人类水平人工智能的普遍追求。

Many of the startup companies and university researchers working in this area believe, like Covariant, that a strategy founded on deep neural networks and reinforcement learning is the best way to fuel progress toward more dexterous robots. One notable exception is Vicarious, a small AI company based in the San Francisco Bay Area. Founded in 2010—two years before the 2012 ImageNet competition brought deep learning to the forefront—Vicarious’s long-term objective is to achieve human-level or artificial general intelligence. In other words, the company is, in a sense, competing directly with higher-profile and far better funded initiatives like those at DeepMind and OpenAI. We’ll delve into the paths being forged by those two companies and the general quest for human-level AI in Chapter 5.

Vicarious 的主要目标之一是构建比典型的深度学习系统更灵活（或者用人工智能研究人员的话来说，更不“脆弱”）的应用程序。这种适应性是任何机器人处理目前由人类承担的各种任务的关键要求。Vicarious 的技术联合创始人 Dileep George 领导了该公司的人工智能研究，他认为，制造能够理解和操纵环境的机器人是实现更通用智能的重要途径，2020 年初，该公司透露，开发面向物流和制造业的多功能机器人将是其近期的主要业务战略。

One of Vicarious’s major objectives has been to build applications that are more flexible—or as AI researchers would say, less “brittle”—than typical deep learning systems. This kind of adaptability is a critical requirement for any robot expected to handle a wide variety of tasks that are currently undertaken by humans. Vicarious’s technical co-founder Dileep George, who has led AI research at the company, believes that building robots with the ability to understand and manipulate their environment is an essential waypoint on the path to achieving more general intelligence, and in early 2020, the company revealed that the development of versatile robots geared toward logistics and manufacturing would be its primary near-term business strategy.

尽管 Vicarious 公司对细节讳莫如深，但它声称已经开发出一种创新的机器学习系统，其灵感来自人脑功能，公司称之为“递归皮层网络”。24^该公司正在将该系统部署到已经投入生产的机器人上，客户包括 Pitney Bowes 的物流部门和化妆品公司 Sephora。Vicarious 的机器人在完成指定任务方面具有显著的改进能力，在首次运行后的数小时内就能取得显著进步。25^其目标是制造出不仅能从货架或箱子里挑选物品的机器人，还要超越这一点，设计出具有真正多功能操作能力的机器，包括分类和包装物品等功能，取代操作工厂机器的工人，送入和取出零件，并执行详细的组装工作。 Vicarious 已筹集至少 1.5 亿美元风险投资，并得到了硅谷一些知名人士的支持，包括埃隆·马斯克、马克·扎克伯格、彼得·泰尔，以及——正如你可能预料的——杰夫·贝佐斯。

Although it is secretive about the details, Vicarious claims to have developed an innovative machine learning system, inspired by the function of the human brain, that it calls a “recursive cortical network.”²⁴ The company is deploying its system to animate robots that have already been placed in production for its initial customers, which include the logistics division of Pitney Bowes and the cosmetic company Sephora. Vicarious’s robots have a remarkable ability to improve at their assigned tasks—getting measurably better within hours of their initial operation.²⁵ The objective is to create robots capable of not just picking items from inventory shelves or bins, but to go beyond that and design machines with genuinely versatile manipulative ability, including functions such as sorting and packing items, replacing the workers who tend factory machines by feeding parts in and out and performing detailed assembly work. Vicarious has raised at least $150 million in venture funding and is backed by some of most prominent names in Silicon Valley, including Elon Musk, Mark Zuckerberg, Peter Thiel and—as you might have expected—Jeff Bezos.

在人工智能领域取得进展的同时，Vicarious 还在推行创新的“机器人即服务”商业模式，该模式最终可能会对一系列行业产生颠覆性影响。Vicarious 并不自行制造或销售机器人，而是从 ABB 等公司购买工业机器人，将其与专有的人工智能软件集成，然后以与临时就业机构安置人工工人大致相同的方式将机器人出租给公司。这样一来，客户公司就不必像使用工业机器人那样进行前期资本投资和长期承诺。这直接解决了使用机器人的最大缺点之一：购买、安装和编程机器的成本很高，因此需要很长时间才能收回投资。然而，传统的工业机器人缺乏人力的灵活性和适应性。每当工厂或仓库内的流程发生变化时（这种情况经常发生，有时在几个月内就会发生变化），就需要对机器人进行耗时且昂贵的重新编程。这是阻碍机器人在这些环境中更广泛部署的主要因素之一。机器人即服务方法与快速训练机器人执行新任务的能力相结合，清楚地表明我们即将迎来一个机器人将与人类工人一样适应的未来。这可能会改变各行各业的游戏规则。

In parallel with its progress in artificial intelligence, Vicarious is also pursuing an innovative “robots as a service” business model, which may eventually prove to be disruptive across a range of industries. Rather than building or selling its own robots, Vicarious instead acquires industrial robots from companies like ABB, integrates them with its proprietary artificial intelligence software and then rents the robots out to companies in a way that’s roughly comparable to the way a temporary employment agency might place human workers. The result is that client companies do not have to make the upfront capital investment and long-term commitment that is normally associated with industrial robots. This directly addresses one of the biggest drawbacks of using robots: the machines are expensive to purchase, install and program, and therefore it takes a long time for the investment to pay for itself. Traditional industrial robots, however, lack the flexibility and adaptability of a human workforce. Anytime the processes within a factory or warehouse change—and this can occur often, sometimes in the space of a few months—a time-consuming and expensive reprogramming of the robots is required. This has been one of the primary factors holding back more widespread deployment of robots in these environments. The robots-as-a-service approach, combined with the ability to rapidly train robots for new tasks, is clear evidence that we are approaching a future where robots will be just as adaptable as human workers. And that is likely to be a game changer in a variety of industries.

Vicarious 并不是唯一一家认识到这种商业模式优势的公司。澳大利亚自动化技术公司 Knapp 也采取了类似的方法，该公司正在使用由 Covariant 软件驱动的机器人。2020 年 1 月，Knapp 高管 Peter Puchwein 告诉《纽约时报》，该公司的策略是始终将机器人的价格定在低于雇用人类工人的成本以下。例如，“如果一家公司每年向一名工人支付 4 万美元，Knapp 就会收取 3 万美元左右的费用。”“我们会把价格定得更低，”Puchwein 告诉《纽约时报》。“这基本上就是我们的商业模式。对于客户来说，决定并不难。” ²⁶当然，除了成本较低之外，机器人还不会休假、永远不会生病、永远不会迟到，而且一般不会遇到人类工人不断出现的任何管理问题和不便。

Vicarious isn’t the only company that has recognized the advantages of this business model. A similar approach is being pursued by the Australian automation technology company Knapp, which is utilizing robots powered by software from Covariant. In January 2020, Knapp executive Peter Puchwein told the New York Times that the company’s strategy is to consistently price its robots below the cost of employing human workers. For example, “if a company paid $40,000 per year to a worker, Knapp would charge about $30,000.” “We just go lower,” Puchwein told the Times. “That’s basically the business model. For the customer, it’s not very hard to decide.”²⁶ Added to the lower cost, of course, is the reality that robots don’t take vacations, never get sick, are never late to work and in general don’t suffer from any of the management issues and inconveniences that arise continuously with human workers.

即使机器人变得更加灵巧，开始接近人类的能力水平，这些机器可能还需要很长时间才能变得足够便宜，可以作为家庭消费品使用。但在工厂和仓库等环境中，事情更加可预测，盈利和效率的逻辑将不可避免地改变工人和机器之间的平衡，这种颠覆可能会来得更快。正如我们所见，机器人不仅在物理操作方面变得更加熟练，而且更加灵活和适应性更强，这将使它们越来越有可能被部署到电子装配等领域，在这些领域，快速转移生产以适应新产品的能力至关重要。所有这些都可能成为人工智能发展成为电力类公用事业的重要篇章，其触角几乎触及经济的方方面面。

Even as robots become vastly more dexterous and begin to approach human levels of capability, it likely will be a long time before these machines become affordable enough to be employed as consumer products in homes. But in environments like factories and warehouses, where things are more predictable and where the logic of profitability and efficiency will inevitably shift the balance between workers and machines, the disruption is likely to come far sooner. As we’ve seen, robots are becoming not just more adept at physical manipulation but also more flexible and adaptable, and this will make them increasingly likely to be deployed even in areas like electronic assembly, where the ability to rapidly shift production to accommodate new products is critical. All this will likely prove to be an important chapter in the story of artificial intelligence’s evolution into an electricity-like utility whose tentacles reach into virtually every aspect of the economy.

最终，对就业的影响将是巨大的，尤其是近年来，随着网上购物继续颠覆传统零售业，仓库和配送中心一直是创造就业的亮点。当这一切强加于正在展开的经济衰退的复苏时，可能会产生特别严重的后果。同样，如果冠状病毒（或者对下一次大流行的挥之不去的恐惧）继续成为一个因素，机器人生产将为围绕社交距离要求或工人生病而出现的问题提供有吸引力的解决方案。我们将在第 6章中更全面地探讨人工智能和机器人技术对就业和经济的潜在影响。

The implications for employment will eventually be significant, especially since warehouses and distribution centers have been a relative job-creation bright spot in recent years as online shopping has continued to disrupt the traditional retail sector. All of this could have especially stark consequences when it is imposed upon recovery from the unfolding economic downturn. Likewise, to the extent that the coronavirus—or for that matter a lingering fear of the next pandemic—continues to be a factor, robotic production will present attractive solutions to problems that arise around a requirement for social distancing or around workers who fall sick. We will explore the potential impact of artificial intelligence and robotics on jobs and the economy more fully in Chapter 6.

传统零售和快餐业即将迎来人工智能革命

THE COMING AI REVOLUTION IN TRADITIONAL RETAIL AND FAST FOOD

2019 年 12 月 3 日，彭博社发表了一篇题为《第二通道的机器人》的文章，深入探讨了美国实体零售商中人工智能、机器人和自动化的兴起。这篇由商业记者马修·博伊尔撰写的文章指出，大型杂货连锁店尤其有兴趣采用新技术，以抵御亚马逊即将进入其市场所带来的潜在生存威胁。古板的杂货行业上一次重大创新是 20 世纪 70 年代末引入条形码扫描仪，现在正紧急试验“货架扫描机器人、动态定价软件、智能购物车、移动结账系统和商店后面的自动迷你仓库”等以人工智能为中心的新技术^。27

On December 3, 2019, Bloomberg published an article entitled “Robots in Aisle Two,” which delved into the rise of artificial intelligence, robotics and automation at brick-and-mortar retailers in the United States. The article, written by business reporter Matthew Boyle, pointed out that the major grocery chains were especially interested in adopting new technology in order to ward off what they perceived as a potentially existential threat from Amazon’s looming entry into their market. The stodgy grocery industry, whose last major innovation was the introduction of bar code scanners in the late 1970s, was now urgently experimenting with “shelf-scanning robots, dynamic pricing software, smart carts, mobile-checkout systems and automated mini-warehouses in the back of stores” among other new AI-centric technologies.²⁷

不过，文章中引用的一位业内人士却发出了缓和的语气。“你不会很快在塔吉特百货看到机器人，”该公司的首席执行官说。“人性化仍然非常重要。” ²⁸大约在这篇文章出现在彭博网站上的两天前，中国武汉出现了第一例有记录的新冠肺炎病例。在接下来的几个月里，我们对“人性化”感知价值的所有计算当然都被重置和重新调整了，速度之快曾经是难以想象的。毫无疑问，在几乎任何人类工人与客户流量直接接触的环境中，冠状病毒危机都将大大加速自动化进程。这不仅是因为人们对社交距离和卫生的担忧，也是因为在病毒引发的经济衰退之后，对效率的关注不可避免地会升级。很有可能，即使当前的危机成为历史（几乎肯定要等到有效的疫苗或治疗方法普及后），这种趋势在很大程度上也是不可逆转的。

Still, an industry insider quoted in the article sounded a moderating note. “You won’t see robots in Target anytime soon,” the company’s CEO said. “The human touch still really matters.”²⁸ Roughly two days before the article appeared on Bloomberg’s website, the first documented case of COVID-19 had emerged in Wuhan, China. Over the course of the next few months, all our calculations around the perceived value of a “human touch” were, of course, reset and recalibrated with a speed that would have once been unimaginable. There can be little doubt that in virtually any environment where human workers come into direct contact with customer traffic, the coronavirus crisis is going to significantly accelerate the push toward automation. This will be true not only because of concerns around social distancing and hygiene, but also as a result of the inevitable escalation of a focus on efficiency in the wake of the economic downturn spawned by the virus. It seems very likely that even when the current crisis passes into history—almost certainly not until an effective vaccine or treatment becomes universally available—this trend could well turn out to be, to a significant extent, irreversible.

从地方杂货店到全国和地区连锁店，各种规模的零售商都在积极部署能够执行专门任务的机器人。例如，自动擦地机器人制造商 Brain Corporation 的销售额大幅增长，因为冠状病毒危机迫切需要商店进行夜间深度清洁。沃尔玛预计到 2020 年底，将在美国 1,800 多家门店安装这些机器。29^这家零售巨头还利用分拣机帮助按部门整理新到货的库存，这些库存是从卡车上卸下来的。同样，零售商也在投资库存扫描机器人，它们会在商店过道里巡逻。沃尔玛计划到 2020 年夏天，在至少 1,000 家门店安装这些高六英尺、配备十五个摄像头的机器，自动检查货架并扫描产品条形码。30^机器人收集的数据被传送到跟踪商店库存的算法中，并立即提醒员工需要补货特定商品。分析显示，缺货商品与店内销售额下降直接相关，因此库存机器人可以立即提高盈利能力，同时为顾客提供更好的体验。事实上，机器学习算法正被用于管理从库存水平到产品选择再到店内特定商品的摆放等所有事情。所有这些都使实体零售商开始利用亚马逊在其在线购物业务中如此有效利用的同类人工智能。

Retailers ranging in size from local grocery stores to national and regional chains have been moving aggressively to deploy robots capable of performing specialized tasks. For example, Brain Corporation, a maker of autonomous floor-scrubbing robots, has seen sales jump dramatically as the coronavirus crisis has put an urgent emphasis on the need for overnight deep cleaning in stores. Walmart expected to place the machines in over 1,800 of its U.S. stores by the end of 2020.²⁹ The retail giant also makes use of sorting machines that help organize newly arrived inventory by department as it is loaded off trucks. Likewise, retailers are investing in inventory-scanning robots that prowl the store aisles. Walmart planned to have the machines, which are six feet tall and equipped with fifteen cameras, autonomously inspect the shelves and scan product barcodes in at least a thousand of its stores by the summer of 2020.³⁰ The data collected by the robots is relayed to algorithms that track store inventory and immediately alert workers of the need to restock particular items. Analysis has shown that out-of-stock items directly correlate with lower in-store sales, so the inventory robots provide an immediate boost to profitability while offering customers a better experience. Indeed, machine learning algorithms are being used to manage everything from inventory levels to product selection to placement of particular items within stores. All this allows physical retailers to begin taking advantage of the same kind of artificial intelligence that Amazon leverages so effectively in its online shopping business.

近期最热门的趋势之一是将所谓的“微型配送中心”整合到传统杂货店的后部。这些设施由多家初创公司建立，包括 Takeoff Technologies 和总部位于以色列的 Fabric，它们提供的机器人配送能力在许多方面可与 Ocado 等公司建造的大型配送中心相媲美。微型配送中心使杂货店能够高效处理在线配送业务，每周最多可准备 4,000 份订单。31通过将在线业务与主店分开，该技术使^杂货店避免派店员进入可能拥挤的过道取货，同时减轻了商店客户区域的库存压力，在冠状病毒卫生纸恐慌时代，这些库存可能已经不足。虽然微型配送中心缺乏规模经济，而大型独立仓库在成本上具有优势，但前期资本支出和将其整合到现有商店所需的时间大大减少——这对小型连锁店或独立商店来说是重要的优势。

One of the hottest recent trends is the integration of so-called “mini–fulfillment centers” into the backs of traditional grocery stores. These facilities, which are set up by a number of startup companies, including Takeoff Technologies and Israeli-based Fabric, offer robotic fulfillment capabilities that are in many ways comparable to what is found in the much larger distribution centers built by companies like Ocado. The mini–fulfillment centers allow grocery stores to efficiently handle online delivery operations and can prepare up to 4,000 orders per week.³¹ By keeping online operations separate from the main store, the technology allows grocers to avoid sending store employees into potentially crowded aisles to retrieve items while reducing pressure on inventory in customer areas of the store which may, in the age of the coronavirus toilet paper panic, already be running low. Though mini–fulfillment centers lack the economies of scale that give larger stand-alone warehouses a cost advantage, the upfront capital outlay and the time required to integrate them into existing stores is substantially reduced—important advantages for smaller chains or independent stores.

总体而言，零售环境中部署的机器人与仓库或工厂中的机器人具有相同的优势和局限性。这些机器可以高效地在商店后部移动和分类材料，并在过道中穿行、擦洗地板或扫描产品条形码。目前，它们无法完成的实际上是将货物放入货架。阻碍机器人革命更广泛应用的根本限制因素往往是灵活性。正如机器人还不能从仓库的货架上挑选各种各样的物品一样，它们还无法胜任将产品放置在零售货架上的更艰巨的工作。当然，随着真正灵巧的机器人开始出现，这种情况注定会改变。

In general, the robots deployed in retail environments exhibit the same strengths and limitations as those found in warehouses or factories. The machines efficiently move and sort materials in the back of the store and navigate the aisles, scrubbing the floors or scanning product barcodes. What they can’t do, for the time being, is actually stock the shelves. The fundamental limitation holding back a more widespread robotic revolution is more often than not dexterity. Just as robots can’t yet pick a wide variety of items from shelves in a warehouse, they are yet not up to the even more exacting job of positioning products on retail shelves. This, of course, is destined to change as genuinely dexterous robots begin to arrive.

值得注意的是，整体零售商业模式正在发生变化。大多数实体店都面临着来自亚马逊和其他在线零售商的持续压力，销售似乎不可避免地将继续逐渐从传统零售环境转向电子商务供应商运营的庞大且自动化程度越来越高的配送中心。即使在食品杂货行业，网上订购和送货也越来越受欢迎，而且由于在冠状病毒危机最严重期间几乎每个人都需要待在家里，这一趋势也大大加速。时间将证明这种消费者偏好的转变是否会是永久性的，但一旦顾客习惯了将食品杂货送到家门口的便利，这种转变似乎将相当持久。这可能会导致零售杂货店的全面重组——商店后部的自动化操作变得相对更重要，而专门用于顾客购物通道的面积和产品库存都会逐渐缩小。最终，我们可能会看到杂货店的出现，它们本质上是仓库，提供几乎即时的送货或取货服务，也许在小区域内，客户可以在通过自助服务终端或移动设备订购之前查看展示的产品。

It’s also important to note that the overall retail business model is changing. Most brick-and-mortar stores are under relentless pressure from Amazon and other online retailers, and it seems inevitable that sales will continue to gradually shift away from traditional retail environments and toward the massive, and ever more automated, distribution centers run by e-commerce providers. Even in the grocery sector, online ordering and delivery is growing in popularity and has been dramatically accelerated by the need for nearly everyone to stay home during the height of the coronavirus crisis. Time will tell whether this shift in consumer preferences proves permanent, but it seems likely that once customers become accustomed to the convenience of having groceries delivered to their doorstep, the transformation will be fairly durable. That could lead to a general restructuring of retail grocery stores—so that the automated operations in the back of the store become relatively more important and both the floor space and product inventory dedicated to customer shopping aisles gradually shrink. Eventually, we might see the emergence of grocery stores that are essentially warehouses offering nearly instant fulfillment for either delivery or pickup, perhaps with small areas where customers can view products on display before ordering through a kiosk or mobile device.

零售自动化的一个特别重要的趋势不需要任何机器人的灵活性，甚至不需要任何移动部件。在一种全新的零售模式——无收银员商店中，购物者只需走进商店，从货架上取下商品，然后离开，无需排队结账、收银，甚至无需明确的付款机制。这一概念最早出现在 2018 年亚马逊的 Go 便利店中。顾客进入这家约 2,000 平方英尺的商店时，首先激活智能手机上的应用程序，然后在经过类似地铁站的旋转门时扫描手机。进入商店后，他们只需从货架上取下商品，直接放入购物袋即可。所有这一切都是由遍布商店天花板的传感器和摄像头的出色组合实现的。虽然亚马逊对细节讳莫如深，但摄像头能够准确地跟踪从货架上取下商品的过程，数据由深度学习系统处理，该系统使用图像识别功能可靠地记录商店中每位顾客在过道中挑选商品时的购买情况。

One especially important trend in retail automation doesn’t require any robotic dexterity or indeed any moving parts at all. In an entirely new retail model—cashierless stores—shoppers simply walk in, grab items from the shelves and then leave without ever encountering a checkout line, cashier or even an explicit payment mechanism. The concept first emerged with Amazon’s Go convenience stores in 2018. Customers enter the roughly 2,000-square-foot stores by first activating an app on their smartphone and then scanning it as they pass through a subway station–like turnstile. Once in the store, they simply remove items from the shelf and place them directly in their shopping bag. All this is enabled by a remarkable synthesis of sensors and cameras clustered on the ceiling throughout the store. While Amazon is secretive about the details, the cameras are able to accurately track products as they are taken from the shelves, and the data is processed by deep learning systems that use image recognition capability to reliably record the purchases of every customer in the store as he or she moves through the aisles selecting items.

这项技术并不完美，确实会发生一些损失，但故意欺骗系统是非常困难的。例如，顾客可以拿走一件商品，然后将其放回货架上，也许是放在不同的位置，然后再次取回，购买仍会正确计算。即使是明显的偷窃行为，例如在拿走商品之前将其遮住或快速将商品放入口袋而不是购物袋中，也很少能得逞。一旦购物者离开商店，再次通过旋转门，购买的商品将自动从顾客的亚马逊账户中扣除^。32

The technology is not perfect, and some losses do occur, but it is remarkably difficult to intentionally deceive the system. Customers can, for example, take an item then put it back on the shelf, perhaps in a different spot, and then retrieve it again, and the purchase will still be correctly tabulated. Even overt attempts to shoplift, such as by obscuring an item before removing it or quickly placing an item in a pocket rather than the shopping bag rarely succeed. Once a shopper leaves the store, again exiting through a turnstile, the purchases are automatically charged to the customer’s Amazon account.³²

亚马逊已在美国主要城市的 26 个地点开设了 Go 便利店，据报道，该公司正在考虑最终在全美开设多达 3,000 家门店。33 ²⁰²⁰年 2 月，该公司宣布开设首家全尺寸无收银员杂货店。这家超市位于西雅图国会山郊区，占地约 10,000 平方英尺，库存约 5,000 种商品。虽然亚马逊一如既往地是最受关注的参与者，但许多初创公司也在竞相将类似技术推向市场。例如，Accel Robotics 在 2019 年 12 月获得了 3000 万美元的风险投资，用于资助其“随手拿货”技术。其他初创公司包括 Trigo、Standard Cognition 和 Grabango，它们都从投资者那里筹集了至少 1000 万美元。34 据^报道，亚马逊还将其技术授权给其他零售商。35^换句话说，我们即将看到一个充满活力且竞争激烈的市场，为无收银通道的商店提供支持。鉴于此，可以肯定的是，各种现有零售商将转向采用新模式。

Amazon has established Go convenience stores in twenty-six locations in major U.S. cities, and by one report is considering eventually opening as many as 3,000 stores across the United States.³³ In February 2020, the company announced its first full-size cashierless grocery store. Located in the Capitol Hill suburb of Seattle, the supermarket is about 10,000 square feet and stocks roughly 5,000 items. While Amazon is, as usual, the highest-profile player, a number of startup companies are racing to bring similar technologies to market. Accel Robotics, for example, received $30 million in venture capital to fund its “grab and go” technology in December 2019. Other startups include Trigo, Standard Cognition and Grabango, all of which have raised at least $10 million from investors.³⁴ Amazon has reportedly also licensed its technology to other retailers.³⁵ In other words, we are on the verge of seeing a vibrant and highly competitive market for the technology that powers stores without checkout aisles. Given this, it’s a good bet that a variety of existing retailers will move toward adopting the new model.

如果无人收银店真的获得发展，就有可能引发行业重大变革，并最终使仅在美国就有 350 多万收银员的工作面临重大风险。除了增加了便利性和无需排队结账节省了时间之外，这些商店可能特别适合受冠状病毒影响的未来，因为它们提供完全无接触支付，无需靠近人工工作人员。讽刺的是，在冠状病毒爆发期间，亚马逊暂时关闭了大部分 Go 商店，也许是因为这些商店非常受欢迎，吸引了大批购物者排队。然而，从长远来看，这项技术似乎非常适合这个至少在一段时间内社交距离至关重要的世界。

If cashierless stores do gain traction, they have the potential to unleash a major industry disruption and eventually put the jobs of more than three and a half million cashiers in the United States alone at significant risk. Beyond the increase in convenience and the time saved by not standing in checkout lines, these stores may be an especially good fit in a future shaped by the coronavirus because they offer completely touch-free payment without the need to ever come in close proximity to a human worker. Ironically, Amazon temporarily shut down most of its Go stores as the coronavirus unfolded, perhaps because the stores are so popular that they attract long lines of shoppers. However, in the long run, the technology seems ideally suited to a world where social distancing is, at least for a time, at a premium.

我认为机器人自动化将在不久的将来产生重大影响的另一个领域是快餐业。例如，麦当劳一直在大力推动在其全球餐厅安装触摸屏点餐亭。据报道，该公司在 2019 年花费了近 10 亿美元购买这些机器，并预计将在其几乎所有美国门店安装它们。36^自动点餐亭在麦当劳的欧洲餐厅已经无处不在。

Another sector where I think robotic automation will have a significant impact in the relatively near future is the fast food industry. McDonald’s, for example, has been undertaking a major push to install touchscreen ordering kiosks in its restaurants worldwide. The company reportedly spent nearly a billion dollars on the machines in 2019 and expected to install them in nearly all its U.S. locations.³⁶ The automated kiosks are already ubiquitous in McDonald’s European restaurants.

在不久的将来，餐厅后勤烹饪和准备食物的工作也可能会变得更加自动化。这些工作已经基本不再需要技术，而是被划分为一系列高度常规的任务。这是行业战略的一部分，旨在保持低工资并适应 2019 年高达 150% 的员工流动率^。37这些工作的机械化性质使得逐步用自动化机器取代工人成为非常可行的。

The jobs in the back of the restaurant cooking and preparing food are also likely to see increased automation in the near future. These jobs have already been largely deskilled and divided into a series of highly routine tasks. This is part of an industry strategy to keep wages low and adapt to employee turnover rates that were as high as 150 percent in 2019.³⁷ The mechanized nature of these jobs makes it highly feasible to gradually substitute automated machines for workers.

目前最成功的例子之一是旧金山的 Creator, Inc。该公司在旧金山市场南区的第一家餐厅里，设计精良、外观美观的机器人每三十秒就能做出一个美食级汉堡。顾客使用移动应用程序定制和订购汉堡。然后，机器人从头到尾完全自动化制作汉堡。整个过程中，没有人接触过食物。而且，这台机器还增加了一些新意，即使在有人类厨师的高端餐厅也找不到。每个汉堡的肉都是现磨的，奶酪也是现磨的；面包是切片的，蔬菜是按订单切的。Creator 的汉堡售价为 6 美元——大约是其他餐厅同等质量汉堡价格的一半。该公司的策略不是制造便宜的机器人汉堡，而是降低劳动力成本，从而加大对食品质量的投资。Creator 将大约 40% 的成本分配给食品，而一家普通餐厅可能会花费 30% ^。38

One of the most successful examples so far is San Francisco–based Creator, Inc. The sophisticated and aesthetically designed robot at the company’s first restaurant in the city’s South of Market area is able to crank out a gourmet-quality hamburger every thirty seconds. Customers customize and order their burger using a mobile app. The robot then completely automates production of the hamburger from start to finish. No human being ever touches the food during the process. And the machine adds twists that you might not find even in high-end restaurants staffed by human cooks. The meat is freshly ground and the cheese freshly grated for each burger; buns are sliced and vegetables are cut to order. Creator sells its burgers for $6—about half of what you might expect to pay for similar quality at other restaurants. The company’s strategy is not to build a cheap robotic hamburger, but rather to reduce labor costs as a means of investing more in food quality. Creator allocates about forty percent of its costs to food, while a typical restaurant might spend thirty percent.³⁸

事实证明，开发和制造一台能够完全自动化生产美食级汉堡的机器并非易事。Creator 成立于 2012 年，我在 2015 年出版的《机器人崛起》一书中写到了这家当时名为 Momentum Machines的公司。在机器人投入生产之前，它花了六年多的时间进行硬件和软件工程、设计和测试，旧金山的工厂于 2018 年 6 月开业。然而，这家获得了谷歌风投和其他硅谷顶级风险投资公司资助的公司现在可能准备迅速扩张，或者将其技术授权给其他餐厅。

It turns out that developing and building a machine capable of fully automating the production of gourmet-quality hamburgers is not a trivial undertaking. Creator was founded in 2012, and I wrote about the company, then called Momentum Machines, in my 2015 book Rise of the Robots. It took more than six years of hardware and software engineering, design and testing before the robot was ready to be put into production, and the San Francisco location was opened in June 2018. However, the company, which has received funding from Google Ventures and other top Silicon Valley venture capital firms, may now be poised to rapidly expand or perhaps license its technology to other restaurants.

Creator 的战略是利用自动化生产高端汉堡包，但很快，其他各种初创公司可能也会加入进来，开发机器人来生产廉价的商品化汉堡包。我认为，最终，大型快餐连锁店以及小型独立餐厅将不可避免地开始引入这些技术。一旦一家大型企业这样做，并能够利用该技术，竞争态势几乎可以保证自动化的广泛应用。

Creator, with its strategy of leveraging automation to produce high-end hamburgers, will likely soon be joined by a variety of other startup companies that are instead developing robots to produce cheap commoditized burgers. Eventually, I think, it’s inevitable that the major fast food chains, as well as smaller independent restaurants, will begin to introduce these technologies. Once one major player does so and is able to capitalize on the technology, a competitive dynamic virtually guarantees widespread automation.

影响也不仅限于汉堡包。企业家们将找到有效的方法，将机器人应用于从披萨到玉米饼再到您最喜欢的咖啡饮料等各种食品的生产。当然，在冠状病毒爆发后，人们普遍认为，在这种环境下，顾客更喜欢与人类员工互动，而不是与机器人互动，这种观点可能会在一定程度上发生颠覆。突然之间，一台可以在完全没有人为接触的情况下制作出完全准备好的食物的机器可能会带来巨大的营销优势。在我写这篇文章的时候，世界各地的餐馆基本上只提供外卖服务。如果随着危机的持续发展，消费者的偏好永久转向外卖用餐，这将进一步削弱人际互动带来的优势，改变餐馆的商业模式和成本结构，并很可能加速整个行业向自动化的过渡。

Nor will the impact be limited to hamburgers. Entrepreneurs will find effective ways to deploy robots in the production of everything from pizza to tacos to your favorite coffee drink. And, of course, the conventional wisdom that customers have a strong preference for interaction with human employees over robots in these kinds of environments could well be turned on its head to some extent in the wake of the coronavirus. Suddenly a machine that can produce fully prepared food with a total absence of human contact may offer significant marketing advantages. As I write this, restaurants across the world have been largely limited to takeout service. In the event that consumer preference undergoes a permanent shift in favor of takeout dining as the crisis continues to unfold, that would tend to further minimize any advantages offered by human interaction, alter the business models and cost structures of restaurants and quite possibly accelerate the transition to automation across the industry.

医疗保健领域的人工智能

ARTIFICIAL INTELLIGENCE IN HEALTHCARE

从 1970 年到 2019 年的半个世纪里，美国医疗支出占 GDP 的比例增长了一倍多，从约 7% 增至约 18%。39其他发达国家医疗支出曲线的上升趋势没有那么极端，目前的支出数字也低于美国，但情况大致相似。例如，在德国、瑞士和英国等国家，同期医疗支出占 GDP 的比例至少翻了一番。40^{这一全球趋势的主要驱动因素是所谓的“成本病”或鲍莫尔效应，这是经济学家威廉·鲍莫尔和威廉·鲍文研究的一种现象，他们在 1966 年的一本专注}^于表演艺术领域成本病的书中对其进行了描述^。41

Over the half-century from 1970 to 2019, healthcare spending as a percentage of gross domestic product in the United States more than doubled, from about seven percent to roughly eighteen percent.³⁹ The upward slope of the healthcare expenditure graph in other developed countries is not as extreme, and the current spending numbers are lower than in the U.S., but the story is broadly similar. In countries including Germany, Switzerland and the United Kingdom, for example, spending as a fraction of GDP has at least doubled over the same period.⁴⁰ The primary driver of this global trend is what’s known as “cost disease” or the Baumol effect, a phenomenon researched by the economists William Baumol and William Bowen, who described it in a 1966 book that focused on cost disease in the performing arts sector.⁴¹

成本病背后的主要思想是，某些经济部门，尤其是医疗保健和高等教育，需要高技能工人进行非常规、不可扩展的努力，因此这些部门的生产力没有像整体经济那样提高。例如，随着工厂自动化的不断发展，单个制造业工人的努力得到了极大的放大。零售和快餐等行业也是如此，新技术的引入以及更高效的工作场所组织、管理技术和商业模式——包括“大型”商店和网上购物的出现——同样提高了生产力。然而，在医疗保健领域，患者仍然需要医生、护士和其他熟练专业人员的高度个性化关注。可以肯定的是，新知识和技术提高了护理质量，并产生了更好的患者治疗效果，但到目前为止，这并没有像我们在工厂工人身上看到的那样放大医疗保健工作者的努力。尽管如此，医疗保健行业的工资必须上涨，以跟上生产率更高的行业工人的收入。如果没有这一点，医生和护士很可能会离开（或永远不会进入）他们的职业，转而寻求其他更有吸引力的机会。结果是医疗保健成本在经济中占据了越来越大的份额^。42

The main idea underlying cost disease is that certain sectors of economy, most notably healthcare and higher education, require non-routine, non-scalable efforts by highly skilled workers, and as a result these sectors have not seen the productivity increases that have manifested in the broader economy. As automation has advanced relentlessly in factories, for example, the efforts of an individual manufacturing worker have been vastly amplified. The same has been true in sectors like retail and fast food, where the introduction of new technology as well as more efficient workplace organization, management techniques and business models—including the advent of “big box” stores and online shopping—have likewise boosted productivity. In healthcare, however, patients continue to require highly individualized attention from doctors, nurses and other skilled professionals. To be sure, new knowledge and technology have increased the quality of care and produced vastly better patient outcomes, but so far, this has not amplified the efforts of healthcare workers in the way that we have seen with factory workers. Nonetheless, wages in the healthcare sector have had to rise to keep pace with what workers in more productive industries earn. Without this, doctors and nurses would be likely to leave (or never enter) their professions in favor of more attractive opportunities elsewhere. The result is that healthcare costs have come to dominate an ever larger share of the economy.⁴²

人工智能面临的最大机遇和挑战之一是找到解决医疗成本问题的方法。人工智能是否能证明它最终能通过扩大整个行业的生产率提高来扭转医疗支出曲线？这还没有发生，但有充分的理由乐观地认为人工智能将在长期内产生重大影响。

One of the greatest opportunities—and challenges—for artificial intelligence is to find a cure for healthcare’s cost disease. Will artificial intelligence prove to be the technology that will finally bend the healthcare spending curve by scaling productivity increases across the industry? It hasn’t happened yet, but there are certainly good reasons to be optimistic that AI will have a significant impact over the long run.

机器人已经在医院中取得了重大进展，但它们也受到我们在仓库和零售环境中看到的相同基本限制。例如，消毒机器人正在迅速普及。这些机器能够创建医院房间的虚拟地图，然后自主导航，同时将强烈的紫外线照射到每个表面。与人类工人不同，机器人从不错过任何一个地方。紫外线可以迅速破坏病毒和细菌中的 RNA 或 DNA，使一个典型的房间在十五分钟左右就可以消毒。事实证明，该程序比液体消毒剂更有效，特别是因为一些最危险的“超级细菌”已经进化出对这些化学物质的抵抗力。一家制造商，总部位于圣安东尼奥的 Xenex，在冠状病毒大流行的前三个月，其消毒机器人的需求增长了 400% ^。43

Robots have already made significant inroads in hospitals, but they are subject to the same basic limitations that we’ve seen in warehouse and retail environments. Disinfecting robots, for example, are rapidly growing in popularity. These machines are able to create a virtual map of a room in a hospital and then autonomously navigate while directing intense ultraviolet radiation at every surface. Unlike a human worker, the robot never misses a spot. The UV light rapidly destroys the RNA or DNA in viruses and bacteria, allowing a typical room to be disinfected in fifteen minutes or so. The procedure has been shown to be significantly more effective then liquid disinfectants, especially since some of the most dangerous “superbugs” have evolved to be resistant to these chemicals. One manufacturer, San Antonio–based Xenex, saw a 400 percent increase in demand for its disinfecting robots in the first three months of the coronavirus pandemic.⁴³

其他机器人在医院的走廊和电梯中自主导航，运送药物、床单和医疗用品。这些机器人能够携带重物，并定期返回充电站充电。同样，大型药房机器人可以完美无缺地准备和分发数千张处方，提高了大型医院的效率并减少了用药错误。这些机器完全自动化了这一过程；从医生在医院的计算机系统中输入订单开始，直到机器人将药物包装并贴上跟踪条形码后，没有人接触过药物。该系统还跟踪药房的库存，并每天自动生成新药订单^。44

Other robots autonomously navigate the hallways and elevators in hospitals, delivering drugs, linens and medical supplies. The robots are able to carry heavy loads and periodically return to charging stations to have their batteries topped up. Likewise, massive pharmacy robots that prepare and dispense thousands of prescriptions with flawless accuracy have increased efficiency and reduced medication errors in major hospitals. The machines completely automate the process; from the time a physician enters the order in a hospital’s computer system, no human touches the medication until after it is packaged and labeled with a tracking barcode by the robot. The system also keeps track of the pharmacy’s inventory and automatically generates orders for new medications on a daily basis.⁴⁴

这些都是重要的进步，但它们仍然局限于医疗环境中最常规的工作。目前还没有机器人能够完成医生和护士所需的高技能干预。像达芬奇系统这样的手术机器人已经非常流行，可能会增强外科医生的能力，但这些机器并不是自主的。相反，现在由原本手动进行手术的医生来操作机器人。患者可能会对结果更满意，但外科医生和随行医疗团队所需的时间并没有大幅减少。医生和护士执行的操作工作对人工智能提出了巨大的挑战，因为它需要极高的灵活性、解决问题和人际交往能力，以及处理不可预测的环境的能力，在这种环境中，每种情况和每个患者都是独一无二的。就物理医疗机器人而言，我们在工厂或仓库中看到的生产力规模效应可能还远未实现，而且不仅需要大幅提高机器人的灵活性，而且很可能需要通用人工智能或非常接近人工智能的东西。

These are important advances, but once again, they’re limited to the most routine aspects of the work that needs to be performed in healthcare environments. There are no robots that can scale the highly skilled interventions required of doctors and nurses. Surgical robots like the da Vinci system have become very popular and may amplify the capabilities of surgeons, but these machines are not autonomous. Instead, the same doctor that would otherwise perform the surgery manually now manipulates the robot. The patient may be happier with the result, but the time required of the surgeon and the accompanying medical team are not dramatically reduced. The manipulative work performed by doctors and nurses presents an extraordinary challenge for artificial intelligence because it requires extreme dexterity combined with problem solving and interpersonal skills, as well as the ability to handle an unpredictable environment where every situation, and every patient, is unique. As far as physical healthcare robots are concerned, the productivity scaling effect that we have seen in factories or warehouses likely lies in the distant future and will require not just vastly improved robotic dexterity, but quite possibly artificial general intelligence or something very close to it.

鉴于物理机器人的局限性，短期内人工智能对医疗保健产生真正重大影响的领域很可能是那些不需要移动部件的活动。换句话说，人工智能将在信息处理和纯智力活动（如诊断或制定治疗计划）中发挥作用。使用机器视觉技术解读医学图像是一个特别有前途的领域。许多研究表明，深度学习系统在许多情况下能够匹敌甚至超越人类放射科医生的能力。例如，谷歌和几所医学院的研究人员团队于 2019 年发表的一项研究表明，深度学习系统能够通过分析 CT 扫描来诊断肺癌，胜过放射科医生。谷歌系统的准确率为 94.4%，在无法获得患者之前的 CT 扫描结果的情况下，“表现优于所有六位放射科医生”，而在有之前的图像可供比较的情况下，“表现与同一位放射科医生相当” ^。45

Given the limitations of physical robots, it seems likely that any truly significant near-term AI impact on healthcare will emerge in activities that require no moving parts. In other words, artificial intelligence will make its mark in the processing of information and in purely intellectual endeavors, such as diagnosis or the development of treatment plans. The interpretation of medical images using machine vision techniques is an especially promising area. A number of studies have demonstrated that deep learning systems are, in many cases, able to match or exceed the capabilities of human radiologists. For example, a study published by a team of researchers from Google and several medical schools in 2019 showed that a deep learning system was able to beat radiologists at diagnosing lung cancer by analyzing CT scans. Google’s system was 94.4 percent accurate and “outperformed all six radiologists” in cases in which a prior CT scan for the patient was not available, and “was on-par with the same radiologists” when a previous image was available for comparison.⁴⁵

同样，由于冠状病毒大流行威胁到医院不堪重负，放射科 AI 系统在某些情况下被紧急使用。在 COVID-19 检测试剂短缺的情况下，胸部 X 光检查显示出通常由病毒引起的肺炎的证据，成为一种重要的替代诊断技术。一些医院出现积压，导致放射科医生难以分析图像，延误了六个小时或更长时间。作为回应，两家 AI 诊断工具制造商——总部位于孟买的 Qure.ai 和韩国公司 Lunit 能够迅速重新校准他们的系统以专注于冠状病毒。一项研究发现，Qure.ai 的系统在区分 COVID-19 和其他引起肺炎的疾病方面准确率高达 95% ^。46

Likewise, radiology AI systems were employed on an emergency basis in some cases as the coronavirus pandemic threatened to overwhelm hospitals. Amidst a shortage of tests for COVID-19, chest X-rays that showed evidence of the pneumonia often caused by the virus became an important alternative diagnostic technique. Some hospitals experienced backlogs resulting in delays of six hours or more as radiologists struggled to analyze the images. In response, two manufacturers of AI diagnostic tools, Mumbai-based Qure.ai and the Korean company Lunit, were able to rapidly recalibrate their systems to focus on the coronavirus. One study found that Qure.ai’s system was ninety-five percent accurate in distinguishing COVID-19 from other conditions that cause pneumonia.⁴⁶

此类成果引发的热情有时会被误认为是炒作。一些深度学习专家几乎理所当然地认为，人工智能系统将在不久的将来完全取代人类放射科医生。图灵奖获得者 Geoffrey Hinton 可以说是深度学习最著名的倡导者，他在 2016 年表示，“我们现在应该停止培训放射科医生”，因为“很明显，五年内深度学习会比放射科医生做得更好。”Hinton 将医生比作卡通人物 Wile E. Coyote，这个人物因经常发现自己“已站在悬崖边上”然后低头一看，最后却跌入深渊而闻名。47^在我写这篇文章时，距离 Hinton 发表声明已经过去了四年，但没有证据表明放射科医生即将失业。事实上，从业者们强烈反驳放射科医生这个职业将很快消失的说法。 2019 年 9 月，斯坦福医学院放射科医生 Alex Bratt 发表了一篇评论文章，题为“为什么放射科医生不必害怕深度学习”，他在文章中指出，基于深度学习的放射系统缺乏灵活性和整体推理能力，通常仅限于处理简单病例。他写道，这些系统无法整合“临床记录、实验室值、先前图像”等信息。因此，该技术迄今为止仅在“无需临床信息或先前研究，仅使用一张图像（或几张连续图像）即可以高特异性和灵敏度检测到的实体”方面表现出色。48^我怀疑 Geoff Hinton 会认为这些限制必然会被克服，从长远来看，他很可能是对的，但我认为这将是一个渐进的过程，而不是突然中断。

Results like these have led to enthusiasm that can sometimes blur into hype, and among some deep learning experts it’s often taken almost as a given that AI systems will completely replace human radiologists in the relatively near future. Turing Award winner Geoffrey Hinton, arguably the most prominent advocate of deep learning, said in 2016 that “we should stop training radiologists now” because “it’s just completely obvious that within five years deep learning is going to do better than radiologists.” Hinton compared the doctors to Wile E. Coyote, the Roadrunner cartoon character famous for often finding himself “already over the edge of the cliff” before looking down and only then plunging into the abyss.⁴⁷ As I write this, four years after Hinton’s statement, there’s no evidence of looming unemployment for radiologists. Indeed, practitioners push back aggressively against the argument that their profession will soon evaporate. In September 2019, Alex Bratt, a doctor in the Department of Radiology at Stanford Medical School, published a commentary entitled “Why Radiologists Have Nothing to Fear from Deep Learning” in which he made the case that deep learning–powered radiology systems lack flexibility and holistic reasoning and are generally limited to simple cases. The systems, he wrote, have no ability to integrate information from “clinical notes, laboratory values, prior images” and the like. As a result, the technology has so far excelled only with “entities that can be detected with high specificity and sensitivity using only one image (or a few contiguous images) without access to clinical information or prior studies.”⁴⁸ I suspect that Geoff Hinton would argue that these limitations will inevitably be overcome, and he will very likely turn out to be right in the long run, but I think it will be a gradual process rather than a sudden disruption.

另一个现实是，除了技术本身的能力之外，还有各种具有挑战性的障碍，这些障碍可能会使放射科医生或其他医疗专家在短期内很难被送上失业大军。医疗保健的几乎每个方面都受到严格监管，有时由多个具有重叠权限的实体监管。将持照医生完全排除在外并非易事。美国医学会等组织的权力使医生对自己的命运拥有比大多数其他类型的工作者更大的影响力。还有重要的责任问题。导致患者不良结果的错误很容易导致医疗事故诉讼。目前，这种责任由数千名医生承担。如果这项工作是由一家财力雄厚的公司开发和销售的设备或算法来完成的，那么责任就会集中起来，并可能引发大量诉讼。这些都是可能在长期内得到解决的问题，但在可预见的未来，我认为问题不是人工智能是否会取代放射科医生，而是它能否显著提高他们的工作效率。如果深度学习可以让放射科医生在给定的时间范围内分析更多的图像，同时提供即时的第二意见以最大程度地降低错误率，那么这将会扩大个体医生的努力，并且随着时间的推移，可能促使医学生根据市场对其服务的自然需求选择不同的专业。

An additional reality is that there are a variety of challenging hurdles beyond the capability of the technology itself that will probably make it very difficult to send radiologists—or any other medical specialists—to the unemployment line anytime soon. Nearly every aspect of healthcare is heavily regulated, sometimes by multiple entities with overlapping authority. Taking licensed physicians completely out of the loop is not going to be easily accomplished. The power of organizations like the American Medical Association gives doctors far more influence over their fate than most other types of workers. There are also important liability issues. An error that leads to a bad outcome for a patient can easily lead to a malpractice lawsuit. Currently this liability is distributed among thousands of individual doctors. If the work is instead performed by a device or algorithm developed and marketed by a deep-pocketed corporation, that would concentrate the liability and potentially create an incentive for a deluge of litigation. These are all issues that may be resolved in the long run, but for the foreseeable future, I think the question is not whether AI will replace radiologists but whether it can significantly boost their productivity. If deep learning allows radiologists to analyze significantly more images over a given time frame while offering an instant second opinion that minimizes the error rate, that will amplify the efforts of individual doctors and may, over time, lead to medical students choosing a different specialty in response to natural market demand for their services.

当然，视觉图像并不是深度学习算法可以获取的唯一信息形式。电子医疗记录的转变产生了大量的数据，这些数据在很多方面都非常适合人工智能的应用。利用这些资源来提高效率、降低成本并改善患者治疗效果可能是人工智能在医疗领域最有前景的短期机会。据一些统计，医疗失误是美国第三大死亡原因，仅次于癌症和心脏病。每年有多达 44 万美国人死于可预防的错误。49 因用药不当或剂量错误而导致的事故尤其普遍^。

Visual images are, of course, not the only form of information accessible to deep learning algorithms. The transition to electronic medical records has generated a massive trove of data that is in many ways ideally suited to the application of artificial intelligence. Leveraging this resource in ways that improve efficiency, cut costs and result in better patient outcomes is probably the single most promising near-term opportunity for AI in healthcare. By some accounts, medical errors are the third leading cause of death in the United States, outstripped only by cancer and heart disease. As many as 440,000 Americans die each year as the result of preventable errors.⁴⁹ Mishaps resulting from the administration of an incorrect medication or the wrong dosage are especially prevalent.

在 2019 年的一项研究中，以色列初创公司 MedAware 的一款 AI 应用程序被投入到波士顿布莱根妇女医院 2012 年至 2013 年期间发生的近 75 万次患者互动的历史数据库。该系统标记了近 11,000 处错误。结果分析表明，MedAware 软件在发现合法错误方面的准确率为 92%，近 80% 的警报提供了有价值的临床信息，并且超过三分之二的这些失误无法通过医院现有的系统识别。除了改善患者治疗效果和挽救潜在生命之外，研究还发现，布莱根妇女医院还可以节省约 130 万美元直接由这些错误导致的治疗费用^。50

In a 2019 study, an AI application from the Israeli startup MedAware was turned loose on a historical database of nearly 750,000 patient interactions that had occurred at Brigham and Women’s Hospital in Boston during 2012 and 2013. The system flagged nearly 11,000 errors. Analysis of the results showed that MedAware’s software was ninety-two percent accurate in uncovering legitimate errors, that nearly eighty percent of the alerts offered valuable clinical information and that more than two thirds of these mishaps would not have been identified with the existing systems in use at the hospital. In addition to improved patient outcomes and potential lives saved, the study found that Brigham and Women’s would have saved about $1.3 million in treatment costs that resulted directly from the errors.⁵⁰

人工智能在患者数据方面的最引人注目的应用之一发生在 2016 年，当时 DeepMind 与英国国家医疗服务体系 (NHS) 签署了一项为期五年的数据共享协议。NHS 向 DeepMind 提供了超过一百万患者的信息。开发的试点应用包括一个可以分析患者记录和测试结果的系统，然后在患者有急性肾损伤危险时立即提醒 NHS 工作人员，以及一个人工智能系统，该系统被证明能够通过医学扫描诊断眼部疾病，其准确度在某些情况下甚至超过了医生。尽管进展令人鼓舞，但当该项目被转移到 DeepMind 的母公司谷歌时，这一安排在 2019 年引发了争议。尽管谷歌声称制定了严格的隐私政策，并且数据经过了仔细的匿名处理，但这家科技巨头有权访问 NHS 患者数据，这一说法立即引起了强烈反对。51^所有这些再次说明了技术本身能力之外的因素（在本例中是感知到的隐私问题）如何显著减缓人工智能在医疗领域的部署。

One of the highest profile applications of artificial intelligence to patient data occurred in 2016 when DeepMind entered into a five-year data-sharing agreement with the U.K.’s National Health Service. The NHS provided DeepMind access to information on over a million patients. The pilot applications developed included a system that could analyze patient records and test results and then instantly alert NHS staff when a patient was in danger of an acute kidney injury, as well as an AI system that proved able to diagnose eye disease from medical scans with an accuracy that in some cases exceeded that of doctors. Though progress was promising, the arrangement exploded into controversy in 2019 when the program was transferred to DeepMind’s parent company, Google. There was an immediate backlash against the specter of the tech giant having access to NHS patient data despite the fact that Google claimed strict privacy policies were in place and the data was carefully anonymized.⁵¹ All this illustrates, once again, how factors beyond the capability of the technology itself—in this case, perceived privacy concerns—can act to significantly slow the deployment of artificial intelligence in the healthcare arena.

人工智能在医疗保健领域取得的一些最令人惊讶的成功发生在心理健康领域。Woebot Labs 是一家成立于 2017 年的硅谷初创公司，它开发了一款聊天机器人，该机器人由类似于 Alexa 和 Siri 中使用的自然语言处理技术驱动，并结合了心理学家开发的精心编写的对话元素。Woebot 的方法本质上是自动化认知行为疗法 (CBT)，这是一种帮助抑郁症或焦虑症患者的行之有效的技术。在聊天机器人发布一周内，已有超过 50,000 人与该应用程序进行了交谈。正如创始人兼首席执行官 Alison Darcy 指出的那样，“如果你惊恐发作，而没有治疗师可以或应该和你同床共枕，Woebot 可以在凌晨 2 点出现。” ⁵²事实上，聊天机器人的 24 小时无限可用（目前免费）在心理健康治疗中是全新的事物，该应用程序已经填补了一个关键空白。即使是在美国拥有医疗保险的工人，也往往无法获得心理健康服务。许多医疗保健系统不达标的发展中国家的情况更加糟糕。在政府难以为民众提供哪怕是最基本的医疗服务的地区，大多数公民几乎不可能获得心理健康专业人士的帮助。Woebot 定期与 130 多个国家的人们交谈，其中许多人通过使用人工智能翻译工具与聊天机器人的英语服务进行交流。53在这个心理健康危机日益明显的世界里，冠状病毒大流行^带来的额外压力和焦虑可能大大加剧了这种危机，对许多人来说，这样的工具可能是唯一可行的解决方案。我认为有点讽刺的是，我们可能自然而然地认为医疗保健这个特定领域最具有人性化，但它也是第一个受益于可扩展的人工智能驱动的生产力改进的领域，我们希望有朝一日这种改进将改变整个行业。

Some of the most surprising successes with artificial intelligence in healthcare are occurring in the mental health arena. Woebot Labs, a Silicon Valley startup founded in 2017, has developed a chatbot powered by natural language processing technology similar to what is used in Alexa and Siri, combined with carefully scripted conversational elements developed by psychologists. Woebot’s approach is essentially to automate cognitive behavioral therapy, or CBT, a proven technique for helping people with depression or anxiety. Within a week of the chatbot’s release, more than 50,000 people conversed with the application. As founder and CEO Alison Darcy points out, “Woebot can be there at 2 a.m. if you’re having a panic attack and no therapist can, or should be, in bed with you.”⁵² Indeed, the chatbot’s unlimited twenty-four-hour availability, currently free of charge, is something entirely new in mental health therapy, and the application is already filling a critical space. Even workers who have health insurance coverage in the United States often have limited access to mental health services. The situation is far worse in many developing countries with substandard healthcare systems. In regions where governments struggle to provide the population with even basic medical care, getting access to a mental health professional is a near impossibility for most citizens. Woebot regularly converses with people in more than 130 countries, many of whom communicate by using AI-powered translation tools to interface with the chatbot’s English-only service.⁵³ In a world where a mental health crisis is becoming increasingly evident, and has likely been greatly exacerbated by the additional stress and anxiety brought on by the coronavirus pandemic, tools like this offer what may be, for many people, the only viable solution. I think it is somewhat ironic that the specific field of healthcare that we might naturally regard as being the most intrinsically human is also the first area to benefit from the kind of scalable AI-driven productivity improvements that we someday hope will transform the industry as a whole.

医疗人工智能领域最重要的可预见和真正具有颠覆性的突破可能是面向一般诊断和治疗的全面可靠系统的出现——换句话说，一种“盒子里的医生”。其目的不是取代医生，而是以一种有效地使最优秀医生的技能和经验民主化的方式增强医生的能力。不难想象，未来强大的诊断人工智能系统将大大提高医生的工作效率，同时创造一种环境，即使是经验不足或水平平庸的医生，也可以与一群精英专家组成的虚拟团队一起与患者接触，为他们提供持续的建议。

The most important foreseeable and genuinely disruptive breakthrough in medical artificial intelligence may turn out to be the advent of a comprehensive and reliable system oriented toward general diagnosis and treatment—in other words, a kind of “doctor in a box.” The point would not be to replace doctors but rather to augment them in a way that effectively democratizes the skill and experience of the very best physicians. It’s easy to imagine a future where a powerful diagnostic AI system dramatically increases the productivity of doctors while creating an environment in which even an inexperienced or mediocre physician navigates patient encounters with what amounts to a virtual team of elite specialists looking over his or her shoulder providing continuous advice.

我们肯定还没有到达那个地步，最早沿着这条道路前进的尝试之一提供了一个警示。在 Watson 在 2011 年的Jeopardy!挑战中获胜后，IBM 积极采取行动，将这项技术重新用于医疗保健和其他行业，并围绕 Watson 建立了一个价值数十亿美元的新业务部门。IBM 的愿景是，Watson 将从各种来源吸收知识；它将从教科书、临床笔记、诊断和基因测试结果以及科学论文中吸收大量信息，然后利用超人的能力将这些点连接起来，即使是最有能力的专家也难以理解。IBM 希望这项技术能够在应用方面带来切实的利益，例如为癌症等复杂疾病制定个性化治疗计划。尽管媒体大肆宣传，并热情洋溢地宣称 Watson“正在上医学院”并准备“对抗癌症” ⁵⁴ ，就好像这是下一场Jeopardy!比赛一样，但至少到目前为止，结果并不令人满意。 2017 年，德克萨斯大学 MD 安德森癌症中心（IBM 最受吹捧的医疗保健合作伙伴之一）在发现 Watson 并没有带来实际好处后终止了与它的合作。55^尽管如此，IBM 仍然对这一想法充满信心，并继续投资，越来越多的其他公司也是如此，其中既有初创公司，也有像谷歌这样的巨头。竞争将继续激烈，因为一项真正成功的技术所带来的投资回报可能是惊人的。我认为成功最终会到来，这是不可避免的，但它可能需要超越当前深度学习方法的人工智能技术——换句话说，需要处于该领域前沿的研究人员正在追求的更通用的智能方面的突破。我们将在第 5 章介绍人工智能前沿正在进行的工作。

We are definitely not there yet, and one of the earliest attempts to move along this path offers a cautionary tale. Immediately after Watson’s triumph in the 2011 Jeopardy! challenge, IBM moved aggressively to repurpose the technology for healthcare and other industries and built a new billion-dollar business unit around Watson. IBM’s vision was that Watson would assimilate knowledge from a wide variety of sources; it would inhale a torrent of information from textbooks, clinical notes, diagnostic and genetic test results and scientific papers, and then leverage a superhuman ability to connect the dots in ways that would elude even the most capable expert. IBM hoped the technology would deliver tangible benefits in applications like the development of personalized treatment plans for complex diseases like cancer. Despite extreme hype and glowing media pieces declaring that Watson was “going to medical school” and preparing to “take on cancer”⁵⁴ as though it were the next Jeopardy! match, the results have, at least so far, been underwhelming. In 2017 the MD Anderson Cancer Center at the University of Texas, one of IBM’s most highly touted healthcare partnerships, discontinued working with Watson after finding no real benefits from the technology.⁵⁵ Still, IBM remains confident and continues to invest in the idea, as do a growing number of other companies including both startups and giants like Google. The competition will continue to be keen as the return on an investment that leads to a truly successful technology is potentially staggering. I think it’s inevitable that success will eventually arrive, but it likely will require AI technologies beyond current approaches in deep learning—or in other words, the kinds of breakthroughs in more general intelligence that researchers at the forefront of the field are pursuing. We’ll cover the work going on at the AI frontier in Chapter 5.

最终，如果真正有能力和强大的系统真的出现，我认为这将为新一类医疗专业人员的出现打开大门。这些人可能是受过本科或硕士学位教育的人，他们接受过专门培训，可以在患者和经批准和监管的医疗人工智能系统之间进行交互。这些低成本的工作人员不会直接取代医生，但可能会在医生的监督下工作，并有能力处理更常规的病例。例如，美国的家庭医生通常会被源源不断的患有相同慢性疾病的患者所淹没，最明显的是肥胖、高血压和糖尿病。与人工智能携手合作的新型从业者可能会大大减轻这种负担，同时扩大地理覆盖范围。美国许多农村地区已经严重缺乏医生，而且随着人口老龄化，这种情况只会恶化。为了解决这些问题，并最终实现最终控制医疗保健成本疾病的生产力增长，我认为我们别无选择，只能更多地依赖医疗机器智能。

Ultimately, if truly capable and robust systems do arrive, I think that could open the door to the emergence of a new class of medical professional. These might be people educated with perhaps an undergraduate or master’s degree and trained specifically to interface between patients and an approved and regulated medical AI system. These lower-cost workers would not directly substitute for physicians but might work under their supervision and have the ability to take on more routine cases. Family doctors in the United States, for example, are generally inundated by a constant stream of patients with the same chronic conditions, most notably obesity, high blood pressure and diabetes. A new class of practitioner working hand in hand with artificial intelligence might go a long way to lessening this burden while also expanding geographic coverage. Many rural areas of the U.S. already have serious shortages of doctors, and this will only worsen as our population ages. To address these issues and eventually achieve the kind of productivity increases that will finally rein in healthcare’s cost disease, I think we will have little choice except to rely far more heavily on medical machine intelligence.

自动驾驶汽车和卡车：比预期更长的等待时间

SELF-DRIVING CARS AND TRUCKS: A LONGER THAN EXPECTED WAIT

埃隆·马斯克承诺到 2020 年底将有 100 万辆机器人出租车上路，这只是自动驾驶汽车行业过度兴奋的最新例证。也许是因为汽车在我们的生活中占据着核心地位，尤其是在美国，没有任何人工智能应用像自动驾驶汽车一样受到如此多的炒作和狂热。自 2004 年和 2005 年国防高级研究计划局 (DARPA) 大挑战赛之后该行业兴起以来，该技术取得了惊人的进步，但同时也经常未能达到过高的预期。2015 年，业内最有见识的人士普遍预测，五年内全自动驾驶汽车将出现在我们的道路上。克里斯·厄姆森 (Chris Urmson) 是该领域的先驱之一，曾担任谷歌自动驾驶汽车子公司 Waymo 的首席技术官，现在是自动驾驶初创公司 Aurora 的首席执行官兼创始人。他曾推测，他当时 11 岁的儿子到 16 岁时可能不需要考驾照。丰田和日产等主要制造商也承诺在 2020 年推出自动驾驶汽车。56^现在，所有这些预测都被推翻了。厄姆森仍然充满信心，他在 2019 年表示，他预计五年内至少有“数百辆”全自动驾驶汽车将在公共道路上部署，⁵⁷十年内可能会有 10,000 辆或更多这样的汽车投入运营。58^我个人认为，即使是这些预测也可能过于乐观。我想说，真正的自动驾驶汽车在未来很多年里都只能是五年后的事情，这是一个真正的危险。

Elon Musk’s promise of a million robotic taxis operating on roads by the end of 2020 is only the most recent example of overexuberance in the autonomous vehicle industry. Perhaps because of the centrality of the automobile to our way of life, especially in the United States, no application of artificial intelligence has been subject to as much hype and hyperbolic enthusiasm as the self-driving car. Since the industry’s emergence following the Defense Advanced Research Projects Agency (DARPA) grand challenges in 2004 and 2005, the technology has achieved astonishing progress while at the same time regularly falling short of overinflated expectations. In 2015, it was widely predicted by the most knowledgeable industry insiders that fully autonomous vehicles would be on our roads within five years. Chris Urmson, one of the pioneers of the field, who was formerly the chief technology officer for Google’s self-driving car spinoff, Waymo, and is now CEO and founder of the autonomous driving startup Aurora, famously speculated that his then-eleven-year-old son might have no need to pursue a driver’s license when he turned sixteen. Major manufacturers including Toyota and Nissan likewise promised self-driving vehicles by 2020.⁵⁶ All those predictions have now been rolled back. Urmson remains confident and said in 2019 that he expects at least “hundreds” of fully autonomous vehicles to be deployed on public roads within five years,⁵⁷ and that there may be 10,000 or more such cars operating within ten years.⁵⁸ My own view is that even those predictions could well turn out to be optimistic. I’d say there’s a real danger that truly autonomous cars are going to remain five years in the future for many years to come.

现实情况是，自动驾驶汽车在高速公路和城市环境中的常规运行（即事情或多或少按预期进行的情况）已基本得到解决。如果公共道路在整体可预测性方面与亚马逊仓库内部相似，那么自动驾驶汽车可能已经得到广泛部署。

The reality is that the routine operation of autonomous cars on both highways and in more urban environments—in other words, situations where things work more or less as expected—has largely been solved. If public roads were anything like the inside of an Amazon warehouse in terms of the overall level of predictability, self-driving cars might already be widely deployed.

当然，问题在于所谓的边缘情况，即几乎无限数量的异常互动和情况，自动驾驶汽车很难或不可能准确预测，或者在许多情况下无法正确解释。大多数自动驾驶汽车计划都依赖于对行驶道路的高度精确的先进地图绘制。因此，意外的道路封闭、施工或交通事故可能会造成问题。恶劣的天气，尤其是大雨或大雪，也会产生重大障碍。但最大的挑战可能是安全地与由不可预测的行人、骑自行车者和司机组成的生态系统互动。在旧金山这样的城市，遇到分心或醉酒的行人并不罕见。即使是那些警觉的人，也常常会做出难以解读的行为，在某些情况下或在某些街区小心翼翼地走下路边，而在其他街区则更为激进。在人口稠密的地区，司机和行人之间的大部分协调都依赖于社交互动，而自动驾驶汽车很难理解或复制这种互动。通过眼神交流、挥手、停下脚步等待司机确认以及许多其他细微行为建立的联系构成了一种无声语言，几乎所有路上的人都能以某种方式理解这种语言。我认为，处理这些类型的互动很可能超出了当今深度学习系统的能力范围。换句话说，真正的自动驾驶汽车可能需要在通用机器智能的道路上走得更远的技术，而这可能需要很长时间。

The problem, of course, is in the so-called edge cases, the virtually infinite number of unusual interactions and situations that are difficult or impossible for a self-driving car to accurately predict or, in many cases, to correctly interpret. Most self-driving car initiatives depend on highly precise advanced mapping of the streets being traveled. Therefore, unexpected road closings, construction or traffic accidents can create problems. Inclement weather, especially heavy rain or snow, also produces major impediments. But the greatest challenge may be to safely interact with an ecosystem populated by unpredictable pedestrians, bicyclists and drivers. In cities like San Francisco, it’s not uncommon to encounter pedestrians who are distracted or drunk. Even those who are alert often act in ways that are a challenge to interpret, stepping tentatively off the curb in some cases, or in certain neighborhoods, and far more aggressively in others. In densely populated areas, much of the coordination between drivers and pedestrians relies on social interactions that would be very difficult for a self-driving car to understand or replicate. A connection achieved through eye contact, a wave of a hand, pausing midstride to wait for a driver’s acknowledgement and numerous other nuanced behaviors make up a kind of unspoken language that is somehow understood by nearly everyone who shares the road. I think it is quite possible that it may turn out that negotiating these types of interactions is simply beyond the capabilities of today’s deep learning systems. In other words, truly autonomous cars may require technology much further along the path toward general machine intelligence, and that could be a long wait.

许多分析师认为，鉴于自动驾驶汽车在城市环境中面临的困难，第一种真正实用的无人驾驶汽车将是长途卡车。毕竟，高速公路上的驾驶问题在很大程度上已经被特斯拉的自动驾驶仪等系统解决了。虽然在高速公路上发生不可预测事件的可能性确实低于在繁忙的城市十字路口，但错误造成的后果会因所涉及的速度和车辆是满载卡车并具有几乎不可估量的动能而大大放大。当然，尽管埃隆·马斯克（Elon Musk）兴致勃勃，但特斯拉的自动驾驶仪系统绝对没有被认证可以在没有专心驾驶员的情况下运行。出于这些原因，我认为我们还需要很长时间才能在公共高速公路上经常看到真正的无人驾驶卡车。

Many analysts believe that, given the difficulties faced by autonomous cars in urban settings, the first truly practical driverless vehicles to appear on our roads will be long-haul trucks. Driving on highways, after all, is a problem that has largely already been solved by systems like Tesla’s autopilot. While it’s certainly true that the likelihood of an unpredictable event is lower on a highway than at a busy urban intersection, the consequences of an error are vastly magnified by the speeds involved and the fact that the vehicle is a fully loaded truck traveling with nearly unfathomable kinetic energy. And, of course, in spite of Elon Musk’s exuberance, Tesla’s autopilot system is in no way certified to operate without an attentive driver at the wheel. For these reasons I think it will be quite a while before we routinely see genuinely unmanned trucks on public highways.

我认为，一家小公司所面临的挑战可能对整个行业都具有重要启示。2017 年初，我受邀参观了旧金山的一家初创公司 Starsky Robotics。该公司的首席执行官兼联合创始人 Stefan Seltz-Axmacher 向我解释了该公司的愿景，即建立一个能够在高速公路上长距离自动驾驶的系统，但卡车由人类操作员通过远程控制进行监控。当车辆离开或接近路线终点，或遇到更复杂的情况时，远程操作员（通常是经过再培训的卡车司机）将通过公司总部类似视频游戏的控制台的蜂窝连接驾驶卡车。Seltz-Axmacher 告诉我，他相信该公司将在未来几年内在美国道路上拥有完全自动驾驶的无人驾驶卡车。尽管我对 Starsky 的团队和他们向我展示的技术印象深刻，但我对他们能否实现这一目标非常怀疑，尤其是考虑到他们需要克服的监管障碍。尽管如此，塞尔茨-阿克斯马赫和他的团队还是超出了我的预期：该公司于 2018 年在一条封闭道路上成功驾驶了一辆无人驾驶卡车，然后在 2019 年成为第一家在公共高速公路上测试全自动卡车（车上没有安全驾驶员）的自动驾驶汽车公司。

I suspect the challenges faced by one small company may contain some important insights for the sector as a whole. In early 2017 I was invited to visit a San Francisco–based startup company called Starsky Robotics. The company’s vision, as explained to me by its CEO and co-founder, Stefan Seltz-Axmacher, was to build a system capable of driving autonomously on highways over long distances but to have the trucks supervised by human operators via remote control. As the vehicles left or approached the endpoints on their route or otherwise encountered more complex situations, the remote operator—generally a retrained truck driver—would drive the truck via a cellular connection from a video game–like console at the company headquarters. Seltz-Axmacher told me he believed the company would have fully autonomous, unmanned trucks on American roads within the next few years. Though I was greatly impressed by Starsky’s team and the technology they showed me, I was very skeptical that they would achieve this, especially given the regulatory hurdles they would need to surmount. Nonetheless, Seltz-Axmacher and his team exceeded my expectations: the company successfully operated a driverless truck on a closed road in 2018, and then in 2019 became the first autonomous vehicle company to test a fully automated truck, with no safety driver on board, on a public highway.

Starsky 还采用了一种非常创新的商业模式。Starsky 并没有直接与越来越多希望开发和授权自动驾驶技术的资金充足的初创公司竞争，而是决定直接进入卡车运输行业，并利用其系统获得竞争优势。该公司的管理层认为，只有将技术开发完全融入卡车运输公司的日常运营中，并利用灵活性仅在合理的情况下部署不断发展的系统，才能取得近期的成功。

Starsky also adopted a very innovative business model. Rather than competing directly with the growing number of well-funded startups hoping to develop and license the technology to enable autonomous driving, Starsky instead decided to directly enter the trucking business and use its system to gain a competitive advantage. The company’s management believed that only by fully integrating development of the technology into the daily operations of a trucking company, and taking advantage of the flexibility to deploy the evolving system only in situations where it made sense, could near-term success be achieved.

遗憾的是，投资者最终没有接受这一愿景，该公司在未能筹集到下一轮所需的风险投资后于 2020 年初被迫关闭。在公司倒闭后撰写的一系列博客文章中，Seltz-Axmacher 指出，深度学习的局限性是阻碍该行业进步的主要挑战之一。“监督式机器学习并没有达到炒作的水平，”他写道，“它不是真正的人工智能”，而是“一种复杂的模式匹配工具”。59^换句话说，一个能够在任何情况下灵活地提供真正自动驾驶的系统，不需要远程人工监督，很可能超出了当今深度学习系统的能力，而且不太可能在不久的将来出现。塞尔茨-阿克斯马赫认为，该行业面临的挑战尚未得到充分重视，投资者错过了在短期内将自动驾驶卡车安全地投入高速公路的机会，部分原因是过于关注完全自动化的承诺和竞争初创公司经常展示的更先进的功能，但还远没有准备好进行实际部署。

Sadly, investors didn’t ultimately buy into this vision, and the company was forced to shut down in early 2020 after failing to raise the next required round of venture capital. In a series of blog posts written after the company’s demise, Seltz-Axmacher pointed to the limitations of deep learning as one of the primary challenges holding back progress in the industry. “Supervised machine learning doesn’t live up to the hype,” he wrote, “it isn’t actual artificial intelligence” but rather “a sophisticated pattern-matching tool.”⁵⁹ In other words, a system with the flexibility to offer truly autonomous driving under all circumstances, without the need for remote human supervision, may well be beyond the capability of today’s deep learning systems and is unlikely to arrive in the near future. Seltz-Axmacher believes that the challenges faced by the industry are not yet fully appreciated and that investors missed an opportunity to safely put a fleet of self-driving trucks on highways in the near term, in part because of an overriding focus on the promise of full automation and on the more advanced features that were often demonstrated by competing startups, but were nowhere close to being ready for real-world deployment.

开发足够强大的技术是自动驾驶汽车行业面临的最大挑战，但我认为，关于此类汽车的潜在商业模式也存在一些真正的问题。部署自动驾驶汽车的自然场所通常被认为是拼车服务。优步及其竞争对手一直在利用风险投资或最近的 IPO 获得的资金来补贴每次乘车的费用。60^鉴于这种做法不可持续，人们普遍认为自动驾驶汽车是长期解决方案。一旦司机（通常获得 70% 到 80% 的车费）退出市场，这些公司应该可以顺利实现盈利。这是优步将自动驾驶汽车公司（尤其是 Waymo）视为生存威胁并选择从 2016 年开始大力投资自己的自动驾驶项目的主要原因。

Developing sufficiently capable technology looms as the greatest challenge for the autonomous vehicle industry, but I think there are also some real questions regarding the potential business models for such vehicles. The natural place to deploy self-driving cars is generally assumed to be in ride-sharing services. Uber and its competitors have been subsidizing the cost of every ride by drawing on capital obtained through venture funding or, more recently, IPOs.⁶⁰ Given that this is unsustainable, self-driving cars are widely viewed as the long-term solution. Once the driver, who generally gets seventy to eighty percent of the fare, is out of the picture, the companies ought to have a smooth path to profitability. This is the primary reason that Uber views autonomous vehicle companies, particularly Waymo, as existential threats and chose to invest heavily in a self-driving program of its own beginning in 2016.

认为自动驾驶技术将拯救 Uber 和 Lyft 的假设存在问题，因为 Uber 和 Lyft 被视为有吸引力的互联网企业，并因此获得相应的估值，因为它们主要充当数字中介，通过提供自动匹配乘客和司机的软件，从每笔交易中分得一杯羹。这使得这些公司完全避免了出租车业务中风险大且令人不快的部分：拥有、融资、维护和投保车辆等。所有这些都被推到了司机身上。Uber 和 Lyft 无需更换机油、洗车或爆胎；他们基本上置身事外，收取干净的互联网费用。然而，摆脱司机也意味着摆脱那些相当方便地拥有和维护汽车的人。一旦汽车实现自动驾驶，这些公司将发现自己拥有大量车辆，因此将负责随之而来的所有麻烦和费用。Uber 实际上将与赫兹或 Avis 非常相似，这两家公司都没有被估值为“科技公司”。此外，由于共享乘车公司需要激光雷达系统等专用设备，其车辆的价格将高得多。在新冠疫情之后，人们可能还会更加重视经常对车辆进行适当的清洁和消毒。这又是司机目前的责任。

The problem with the assumption that self-driving technology will ride to their rescue is that Uber and Lyft are viewed as attractive internet-based businesses—and valued accordingly—because they act primarily as digital intermediaries, harvesting a slice of every transaction in return for providing software that automatically matches riders with drivers. This allows the companies to completely avoid the risky and unpleasant parts of the taxi business: stuff like owning, financing, maintaining and insuring vehicles. All of that gets pushed onto the drivers. No oil changes, car washes or flat tires for Uber or Lyft; they largely remain above the fray, hoovering up clean internet fees. Getting rid of the drivers, however, also means getting rid of the people who, rather conveniently, own and maintain the cars. Once the cars become autonomous, the companies will find themselves in the business of owning vast fleets of vehicles and will therefore be responsible for all the hassle and expense that comes along with that. Uber in effect will look quite a lot like Hertz or Avis—neither of which is valued as a “tech company.” Moreover, the vehicles owned by the ridesharing companies will be far more expensive given the specialized equipment, such as lidar systems, that they require. In the aftermath of the coronavirus pandemic, there may also be far more emphasis placed on properly cleaning and sanitizing vehicles on a frequent basis. This, again, is something that is currently the responsibility of drivers.

我认为，观察未来几年自动驾驶汽车的发展将会非常有趣，无论是技术还是最终出现的商业模式。硅谷有大量初创公司专注于开发和授权自动驾驶技术，几乎每家大型汽车制造商都在不同程度上进行投资。这些举措中的任何一项都可能带来颠覆性的突破，但我认为最有趣的故事之一将集中在 Waymo 和特斯拉所采取的战略之间的差距越来越大，以及这两家公司之间的竞争将如何随着时间的推移而展开。

I think it will be fascinating to watch the evolution of self-driving cars over the coming years in terms of both the technology and the business models that ultimately emerge. There are a large number of Silicon Valley startups focused on developing and licensing self-driving technology, as well as varying degrees of investment by virtually every major automotive manufacturer. A disruptive breakthrough could emerge from any of these initiatives, but I think one of the most interesting narratives will center on the widening gap between the strategies pursued by Waymo and Tesla, and how competition between these two companies plays out over time.

Waymo 是谷歌 2009 年启动的自动驾驶汽车项目的直系后代，它拥有比其他任何人都多的经验，被普遍视为行业领导者。Waymo 是唯一一家提供自动驾驶汽车运营服务的公司，付费客户已经可以乘坐无人驾驶汽车。这项名为 Waymo One 的服务目前仅适用于凤凰城郊区精心绘制的（或“地理围栏”）区域的预定路线。那里的道路很宽阔，天气也很好，行人也很少。换句话说，这项服务与在旧金山或曼哈顿叫一辆 Uber 去你想去的任何地方都相去甚远。尽管如此，Waymo One 还是一项令人印象深刻的壮举，我认为它或多或少就是可预见的未来自动驾驶汽车服务的样子：在精心策划的不太具有挑战性的地区指定路线并指定停靠点。当然，这再次引发了一个问题：如此有限的运营如何才能盈利。完全自动驾驶汽车（在非常昂贵的车辆上）要便宜到什么程度，才能让顾客选择它，而不是由人工驾驶的 Uber 或 Lyft 提供的更加灵活的门到门服务？

Waymo, the direct descendant of Google’s self-driving car program initiated in 2009, has more experience than anyone else and is generally regarded as the industry leader. Waymo is the only company that offers an operating automated car service with which paying customers can already ride in a car with no driver at the wheel. This service, called Waymo One, is currently available only for predefined routes in a carefully mapped—or “geo-fenced”—region of suburban Phoenix. The roads are wide, the weather is predictably cooperative and pedestrians are sparse. In other words, the service is a far cry from hailing an Uber and going wherever you like in San Francisco or Manhattan. Nonetheless, Waymo One is an impressive feat, and I think it is more or less what self-driving car service will look like for the foreseeable future: specified routes with designated stops in carefully curated areas that aren’t too challenging. Of course, this once again raises the question of how such a limited operation can become profitable. How inexpensive does a fully automated ride (in a very expensive vehicle) have to be to get a customer to opt for it over the far more flexible door-to-door service offered by a human-driven Uber or Lyft?

Waymo 谨慎行事，值得称赞；而特斯拉则不断突破极限，经常进入业内许多人认为近乎鲁莽的领域。该公司告诉现有客户，他们的汽车拥有支持完全自动驾驶所需的所有硬件，最终将通过软件更新实现该功能。这是一个非常雄心勃勃的承诺。特斯拉还与 Waymo 和业内几乎所有其他公司不同，它放弃了激光雷达——一种通过发射激光然后检测反射光来跟踪汽车周围物体的系统。激光雷达价格昂贵，而且至少在目前的实例中，外观丑陋。特斯拉独特地认为，仅依靠摄像头和雷达就可以实现完全自动化。正如我之前提到的，特斯拉在其汽车上的多个摄像头收集的数据方面享有显著优势。Waymo 拥有约 600 辆自动驾驶汽车。特斯拉拥有一支不断扩大的车队，拥有超过 40 万辆汽车在路上收集数据。 Waymo 的汽车在实际道路上行驶了数百万英里，在模拟中行驶了数十亿英里。61^特斯拉的汽车在自动驾驶系统的控制下行驶了数十亿英里。所有这些在实际道路上收集的数据显然是一个优势，但最终的成功将取决于足够强大的人工智能来利用这些资源，我认为，当今的深度学习技术是否能胜任这项任务是一个真正的问题。

While Waymo proceeds deliberately and with laudable caution, Tesla, in contrast, continuously pushes the envelope, often transgressing into territory that many in the industry feel borders on reckless. The company has told its existing customers that their cars have all the necessary hardware to support fully autonomous driving, and that the capability will eventually be enabled through a software update. This is an extraordinarily ambitious promise. Tesla has also diverged from Waymo and virtually everyone else in the industry by forgoing lidar—systems that track objects around the car by firing a laser and then detecting the reflected light. Lidar is expensive and, at least in its current instantiation, ugly. Tesla uniquely believes that it can achieve full automation by relying solely on cameras and radar. As I noted previously, Tesla enjoys a significant advantage in terms of the data the multiple cameras on its cars collect. Waymo has a fleet of about 600 self-driving vehicles. Tesla has an expanding fleet of over 400,000 cars on the road collecting data. Waymo’s vehicles have driven millions of miles on actual roads and billions of miles in simulation.⁶¹ Tesla’s cars have driven billions of real-world miles while operating under the control of its autopilot system. All this data collected on actual roads is a clear advantage, but ultimately success will depend on artificial intelligence that is sufficiently powerful to leverage that resource, and I think there are real questions as to whether today’s deep learning technology is up to that task.

业界面临的另一个重要问题是最终将提供的自动驾驶水平。自动驾驶系统分为五类。1 至 3 级表示系统本质上是辅助性的。汽车可以在有限的情况下自动驾驶，例如在高速公路上巡航时，但驾驶员必须保持警惕，随时准备控制汽车。包括特斯拉在内的大多数汽车制造商都专注于提供这一范围内的功能。问题在于，由于该系统几乎一直正常工作，驾驶员不可避免地会分心。例如，许多特斯拉司机告诉我，他们经常在硅谷高速公路的拼车车道上使用汽车的自动驾驶系统时用手机回复电子邮件。这种行为已经导致了致命的事故。目前还不清楚汽车如何在长时间的常规驾驶中成功地强制驾驶员集中注意力。自动驾驶系统最大的卖点之一是，它承诺未来有一天将大幅减少全球每年死于交通事故的人数（超过 130 万）。62^如果仅起到辅助作用的系统本身仍存在危险，那么它们可能不足以对这一数字产生有意义的影响。

Another important question for the industry surrounds the level of autonomy that will ultimately be provided. Autonomous driving systems are divided into five categories. Levels 1 through 3 designate systems that are assistive in nature. The car can drive itself under limited circumstances, for example while cruising on a highway, but the driver must remain alert and ready to take control of the car at a moment’s notice. Most automotive manufacturers, including Tesla, are focused on providing capability in this range. The problem is that because the system will work correctly virtually all the time, drivers will inevitably be lured into inattention. A number of Tesla drivers have told me, for example, that they routinely answer email on their phones while using their car’s autopilot system in the carpool lanes of Silicon Valley’s freeways. This kind of behavior has already led to fatal accidents. It’s unclear how the car can successfully enforce attentiveness on the part of the driver over long stretches of routine driving. One of the strongest selling points for self-driving systems is the promise that they will one day dramatically reduce the huge number of people—more than 1.3 million globally—who die each year in traffic accidents.⁶² If systems that are merely assistive continue to come with dangers of their own, they may not be sufficient to put a meaningful dent in this number.

出于这个原因，Waymo 和该领域的许多其他小型初创公司决定专注于 4 级和 5 级自动驾驶。这意味着您可以在其中入睡。事实上，它可能没有刹车踏板或方向盘。在这方面，特斯拉再次成为一个戏剧性的异类。该公司声称它可以弥合这两个愿景之间的差距，通过软件更新可以立即将其汽车从 2 级升级到 4 级自动驾驶，这至少可以说是了不起的。许多人可能会说，这是一个完全夸张的承诺，只不过是空谈。如果特斯拉能在短期内实现这一目标，我会感到惊讶，但如果该公司能够做到这一点，我认为它将把自己定位为明显的行业领导者。事实上，这种预期可能在某种程度上已经计入了公司股价。

For this reason, Waymo, along with many of the other smaller startups in this space, have made a decision to focus exclusively on level 4 and 5 autonomy. This indicates a self-driving car in which you can go to sleep. Indeed, it may not have a brake pedal or a steering wheel. Here again, Tesla is a dramatic outlier. The company’s claim that it can bridge the gap between these two visions, with a software update that will instantly upgrade its cars from level 2 to level 4 autonomy is, to say the least, remarkable. Many might say it’s a promise that is completely over the top and little more than vaporware. I will be astonished if Tesla can achieve this anytime soon, but if the company can do it at all, it will, I think, position itself as the clear industry leader. Indeed, that expectation may to some extent already be factored into the price of the company’s stock.

埃隆·马斯克和特斯拉管理团队的其他成员显然对完全自动驾驶的前景进行了深思熟虑。除了技术之外，他们还开发了解决商业模式问题的潜在方案。在 2019 年的自动驾驶日活动上，马斯克描述了一项计划，特斯拉车主可以让他们的汽车参与该公司运营的自动驾驶出租车服务。特斯拉将从拼车费中分一杯羹，就像苹果从其应用商店中赚取收入一样。这项提议的一个有趣之处在于，它解决了可能最终困扰 Uber 和 Lyft 等公司的所有权和维护问题。特斯拉可能已经找到了一种方法，可以扮演纯粹的互联网中介的角色，同时避免拥有一支汽车车队的需要。大多数特斯拉车主可能不想与陌生人分享他们的汽车，但如果该计划被证明可行，许多客户可能会将特斯拉的汽车作为商业投资，而不是作为私家车购买。

Elon Musk and the rest of Tesla’s management team have clearly given a lot of thought to the prospects for full autonomy. Aside from the technology, they’ve also developed a potential solution to the business model problem. At the 2019 Autonomy Day event, Musk described a scheme in which Tesla owners would be able to have their cars participate in a robotaxi service run by the company. Tesla would get a cut of the ride-sharing fee in the same way that Apple generates revenue from its App Store. One interesting thing about this proposal is that it solves the ownership and maintenance problem that might eventually plague companies like Uber and Lyft. Tesla may have found a way to step into the role of a pure internet intermediary, while avoiding the need to own a fleet of cars. Most Tesla owners might not want to share their vehicles with strangers, but if the plan proves viable, many customers would presumably buy Tesla’s vehicles as a business investment, rather than as personal cars.

毫无疑问，自动驾驶汽车终有一天会成为人工智能革命最实在、影响最深远的体现之一。这项技术有可能重塑我们的城市和生活方式，同时拯救成千上万的生命。然而，我认为我们需要等待十年甚至更长时间，这项技术才能真正到来。人工智能革命的有力证据将首先出现在其他领域——仓库、办公室和零售店等地方——这些领域的技术挑战更易于管理，环境更易于控制，技术不受政府监管，错误造成的后果远没有那么严重。然而，想象一下特斯拉的一次软件更新就能证明我错了，这真是令人兴奋。

There is little doubt that self-driving vehicles will someday be one of the most tangible and consequential manifestations of the revolution in artificial intelligence. The technology has the potential to reshape both our cities and our way of life while saving many thousands of lives. However, I think we will need to wait a decade or more before the technology really arrives. Strong evidence of the AI revolution will first emerge in other areas—places like warehouses, offices and retail stores—where the technical challenges are more manageable, the environment is more controllable, the technology is less subject to government regulation and the consequences of an error are far less dire. It is very exciting to imagine, however, that a single software update from Tesla could prove me wrong.

冲破创新高原：科学和医学研究

BLASTING OFF THE INNOVATION PLATEAU: SCIENTIFIC AND MEDICAL RESEARCH

在那些可能被称为“技术乐观主义者”的人中，人们认为我们生活在一个技术飞速发展的时代。他们告诉我们，创新的速度是前所未有的，呈指数级增长。最热情的加速论者——通常是雷·库兹韦尔的追随者，后者在“加速回报定律”中阐述了这一理念——相信在未来一百年里，按照历史标准，我们将经历“更接近 20,000 年的进步” ^。63

Among those who might be described as “technoptimists,” it is taken as a given that we live in an age of startling technological acceleration. The pace of innovation, we are told, is unprecedented and exponential. The most enthusiastic accelerationists—often acolytes of Ray Kurzweil, who codified the idea in his “Law of Accelerating Returns”—are confident that in the next hundred years, we will experience, by historical standards, the equivalent of something “more like 20,000 years of progress.”⁶³

然而，仔细观察就会发现，虽然加速是真的，但这种非凡的进步几乎只局限于信息和通信技术领域。指数级增长的叙述实际上是摩尔定律及其所实现的越来越强大的软件的故事。在这个领域之外，在由原子而不是比特组成的世界里，过去半个世纪左右的故事截然不同。交通、能源、住房、公共基础设施和农业等领域的创新速度不仅远远达不到指数级增长，甚至可能更确切地说是停滞不前。

Closer scrutiny, however, reveals that while the acceleration has been real, this extraordinary progress has been confined almost exclusively to the information and communications technology arena. The exponential narrative has really been the story of Moore’s Law and the ever more capable software it makes possible. Outside this sector, in the world composed of atoms rather than bits, the story over the past half-century or so has been starkly different. The pace of innovation in areas like transportation, energy, housing, physical public infrastructure and agriculture not only falls far short of exponential, it might be better described as stagnant.

如果你想想象一种以不断创新为特征的生活，想象一下出生于 19 世纪末，然后经历了 20 世纪 50 年代或 60 年代的人。这样的人将见证整个社会几乎难以想象的系统性变革：提供清洁水和处理城市污水的基础设施；汽车、飞机、喷气推进器以及随后的航天时代的到来；电气化以及后来成为可能的照明、收音机、电视和家用电器；抗生素和批量生产的疫苗；美国人的预期寿命从不到 50 岁增加到近 70 岁。相比之下，出生于 20 世纪 60 年代的人将见证个人电脑和后来互联网的兴起，但几乎所有其他在过去几十年中具有如此彻底变革性的创新充其量也只是取得了渐进式的进步。您今天驾驶的汽车与 1950 年的汽车之间的差异，根本无法与 1950 年的汽车与 1890 年的交通方式之间的差异相比。同样，遍布现代生活的几乎每个方面的无数其他技术也是如此。

If you want to imagine a life defined by relentless innovation, think of someone born in the late 1800s who then lived through the 1950s or 1960s. Such a person would have seen systemic transformations across society on an almost unimaginable scale: infrastructure to deliver clean water and manage sewage in cities; the automobile, the airplane, jet propulsion and then the advent of the space age; electrification and the lighting, radios, televisions, and home appliances it later made possible; antibiotics and mass-produced vaccines; an increase in life expectancy in the United States from less than 50 years to nearly 70. A person born in the 1960s, in contrast, will have witnessed the rise of the personal computer and later the internet, but nearly all the other innovations that had been so utterly transformative in previous decades would have seen at best incremental progress. The difference between the car you drive today and the car that was available in 1950 simply does not compare to the difference between that 1950 automobile and the transportation options in 1890. And the same is true of a myriad of other technologies distributed across virtually every aspect of modern life.

计算机和互联网领域的所有显著进步本身并不足以满足人们对前几十年那种广泛进步持续不减的期望，彼得·泰尔（Peter Thiel）有句名言：“我们被承诺会飞的汽车，结果却只得到了 140 个字符。”经济学家泰勒·考恩（Tyler Cowen）和罗伯特·戈登（Robert Gordon）都曾详细阐述过我们一直生活在一个相对停滞的时代——即使信息技术一直在加速发展。泰勒·考恩于 2011 年出版了《大停滞》一书， ⁶⁴而罗伯特·戈登则在 2016 年出版的《美国增长的兴衰^》一书中为美国描绘了非常悲观的未来。65这两本书的一个关键论点是，技术创新的低垂果实在 20 世纪 70 年代左右已基本收获。结果是，我们现在处于技术低谷，人们都在努力攀爬创新树的更高枝。考恩乐观地认为，我们最终将摆脱技术瓶颈。戈登则不那么认为，他认为甚至树的上部树枝也可能已被剥光，我们最伟大的发明可能已经成为过去。

The fact that all the remarkable progress in computing and the internet does not, by itself, measure up to the expectation that the kind of broad-based progress seen in earlier decades would continue unabated is captured in Peter Thiel’s famous quip that “we were promised flying cars and instead we got 140 characters.” The argument that we have been living in an age of relative stagnation—even as information technology has continued to accelerate—has been articulated at length by the economists Tyler Cowen, who published his book The Great Stagnation in 2011,⁶⁴ and Robert Gordon, who sketches out a very pessimistic future for the United States in his 2016 book The Rise and Fall of American Growth.⁶⁵ A key argument in both books is that the low-hanging fruit of technological innovation had been largely harvested by roughly the 1970s. The result is that we are now in a technological lull defined by a struggle to reach the higher branches of the innovation tree. Cowen is optimistic that we will eventually break free of our technological plateau. Gordon is much less so, suggesting that even the upper branches of the tree are perhaps denuded and that our greatest inventions may be behind us.

虽然我认为戈登过于悲观，但有大量证据表明，新创意产生的普遍停滞是真实存在的。2020 年 4 月，斯坦福大学和麻省理工学院的经济学家团队发表了一篇学术论文，发现各行各业的研究效率都急剧下降。他们的分析发现，美国研究人员产生创新的效率“每 13 年下降一半”，换句话说，“为了维持人均 GDP 的持续增长，美国必须每 13 年将研究工作量翻一番，以抵消寻找新创意的难度增加。” ⁶⁶经济学家写道：“无论我们放眼何处，我们都会发现创意以及它们所暗示的指数增长越来越难找到。” ⁶⁷值得注意的是，这种情况甚至延伸到了一直持续产生指数级进步的一个领域。研究人员发现，“摩尔定律所暗示的计算机芯片密度翻倍，如今所需的研究人员数量是 20 世纪 70 年代初所需数量的 18 倍以上”。68^对此的一个可能解释是，在突破研究前沿之前，你必须先了解最新技术。在几乎每个科学领域，这都需要吸收比以前多得多的知识。结果是，创新现在需要由具有高度专业背景的研究人员组成的更大团队，而协调他们的努力本质上比小团队更困难。

While I think Gordon is far too pessimistic, there is plenty of evidence to suggest that a broad-based stagnation in the generation of new ideas is quite real. An academic paper published in April 2020 by a team of economists from Stanford and MIT found that, across a variety of industries, research productivity has sharply declined. Their analysis found that the efficiency with which American researchers generate innovations “falls by half every 13 years,” or in other words “just to sustain constant growth in GDP per person, the United States must double the amount of research effort every 13 years to offset the increased difficulty of finding new ideas.”⁶⁶ “Everywhere we look,” wrote the economists, “we find that ideas, and the exponential growth they imply, are getting harder to find.”⁶⁷ Notably this extends even to the one area that has continued to generate consistent exponential progress. The researchers found that the “number of researchers required today to achieve the famous doubling of computer chip density” implied by Moore’s Law “is more than 18 times larger than the number required in the early 1970s.”⁶⁸ One likely explanation for this is that before you can push through the research frontier, you first have to understand the state of the art. In virtually every scientific field, that requires the assimilation of vastly more knowledge than has been the case previously. The result is that innovation now demands ever larger teams made up of researchers with highly specialized backgrounds, and coordinating their efforts is inherently more difficult than would be the case with a smaller group.

可以肯定的是，还有许多其他重要因素可能导致创新放缓。物理定律规定，可获得的创新并非均匀分布在各个领域。当然，航空航天工程没有摩尔定律。在许多领域，要达到下一个创新成果集群可能需要巨大的飞跃。过度或无效的政府监管当然也发挥了一定作用，目前企业界盛行的短期主义也是如此。长期的研发投资往往与对季度收益报告的过度关注或短期股票表现与高管薪酬的结合不相容。尽管如此，在某种程度上，应对日益复杂的需求和知识爆炸式增长阻碍了创新的步伐，人工智能很可能被证明是我们可以利用的最强大的工具，以摆脱技术瓶颈。我认为这是人工智能在继续发展成为一种无处不在的实用工具时最重要的机会。从长远来看，就我们的持续繁荣和应对已知和意外挑战的能力而言，没有什么比增强我们的集体创新和构思新想法的能力更重要。

To be sure, there are many other important factors that might be contributing to the slowdown in innovation. The laws of physics dictate that accessible innovations are not distributed homogeneously across fields. There is, of course, no Moore’s Law for aerospace engineering. In many areas, reaching the next cluster of innovation fruit may require a giant leap. Over- or ineffective government regulation certainly also plays a role, as does the short-termism that now prevails in the corporate world. Long-term investments in R&D are often not compatible with an obsessive focus on quarterly earnings reports or the coupling of short-term stock performance and executive compensation. Still, to the extent that the need to navigate increased complexity and an explosion of knowledge is holding back the pace of innovation, artificial intelligence may well prove to be the most powerful tool we can leverage to escape our technological plateau. This, I think, is the single most important opportunity for AI as it continues to evolve into a ubiquitous utility. In the long run, in terms of our sustained prosperity and our ability to address both the known and unexpected challenges that lie before us, nothing is more vital than amplifying our collective ability to innovate and conceive new ideas.

人工智能，尤其是深度学习，在科学研究中最有前景的近期应用可能是发现新的化学化合物。正如 DeepMind 的 AlphaGo 系统面对的是一个几乎无限的游戏空间（围棋棋盘的可能配置数量超过了宇宙中的原子数量），涵盖所有可能的分子排列的“化学空间”在实际应用中同样是无限的。在这个空间中寻找有用的分子需要进行极其复杂的多维搜索。需要考虑的因素包括分子结构的三维大小和形状，以及极性、溶解度和毒性等许多其他相关参数。69 对于^化学家或材料科学家来说，筛选替代方案是一个劳动密集型的实验反复试验过程。寻找真正有用的新化学品很容易耗费大量的职业生涯。例如，如今在我们的设备和电动汽车中随处可见的锂离子电池，其研究始于 20 世纪 70 年代，但直到 20 世纪 90 年代才开始商业化。人工智能有望大大加快这一进程。从很多方面来看，寻找新分子非常适合深度学习；算法可以根据已知有用的分子特征进行训练，或者在某些情况下，根据控制分子结构和相互作用的规则进行训练^。70

The most promising near-term application of artificial intelligence, and especially deep learning, in scientific research may be in the discovery of new chemical compounds. Just as DeepMind’s AlphaGo system confronts a virtually infinite game space—where the number of possible configurations of the Go board exceeds the number of atoms in the universe—“chemical space,” which encompasses every conceivable molecular arrangement, is likewise, for practical purposes, infinite. Seeking useful molecules within this space requires a multi-dimensional search of staggering complexity. Factors that need to be considered include the three-dimensional size and shape of the molecular structure as well as numerous other relevant parameters like polarity, solubility and toxicity.⁶⁹ For a chemist or materials scientist, sifting through the alternatives is a labor-intensive process of experimental trial and error. Finding a truly useful new chemical can easily consume much of a career. The lithium-ion batteries that are ubiquitous in our devices and electric cars today, for example, emerged from research that was initiated in the 1970s but produced a technology that could begin to be commercialized only in the 1990s. Artificial intelligence offers the promise of a vastly accelerated process. The search for new molecules is, in many ways, ideally suited to deep learning; algorithms can be trained on the characteristics of molecules known to be useful, or in some cases on the rules that govern molecular configuration and interaction.⁷⁰

乍一看，这似乎是一个相对狭窄的应用。然而，寻找有用的新化学物质的探索几乎涉及创新的每一个领域。加速这一进程有望带来用于机器和基础设施的创新型高强度材料、用于更好的电池和光电管的反应性物质、可能减少污染的过滤器或吸收材料，以及一系列可能彻底改变医学的新药。

At first blush, this may seem like a relatively narrow application. However, the quest to find useful new chemical substances touches virtually every sphere of innovation. Accelerating this process promises innovative high-tensile materials for use in machines and infrastructure, reactive substances to be deployed in better batteries and photoelectric cells, filters or absorbent materials that might reduce pollution and a range of new drugs with the potential to revolutionize medicine.

大学研究实验室和越来越多的初创公司都对机器学习技术充满热情，并已开始使用强大的基于人工智能的方法来取得重大突破。2019 年 10 月，荷兰代尔夫特理工大学的科学家宣布，他们能够完全依靠机器学习算法设计出一种全新的材料，而无需进行任何实际的实验室实验。这种新物质坚固耐用，而且如果对其施加超过一定阈值的力，它还具有超强的可压缩性。这意味着这种材料可以有效地被挤压到其原始体积的一小部分。据该项目首席研究员之一米格尔·贝萨 (Miguel Bessa) 称，具有这些特性的未来材料有朝一日可能意味着“自行车、餐桌和雨伞等日常物品可以折叠起来放进口袋。” ⁷¹

Both university research labs and an expanding number of startup companies have turned to machine learning technology with enthusiasm and are already using powerful AI-based approaches to generate important breakthroughs. In October 2019, scientists at Delft University of Technology in the Netherlands announced that they were able to design a completely new material by exclusively relying on a machine learning algorithm, without any need for actual laboratory experiments. The new substance is strong and durable but also super-compressible if a force beyond a certain threshold is exerted on it. This implies that the material can effectively be squeezed into a small fraction of its original volume. According to Miguel Bessa, one of the lead researchers on the project, futuristic materials with these properties might someday mean that “everyday objects such as bicycles, dinner tables and umbrellas could be folded into your pocket.”⁷¹

此类计划通常要求研究人员拥有强大的人工智能技术背景，但其他大学的团队正在开发更易于使用的基于人工智能的工具，这些工具有望推动新化合物的发现。例如，康奈尔大学的研究人员正在开展一个名为 SARA（科学自主推理代理）的项目，该团队希望该项目能够“以数量级的速度显著加快新材料的发现和开发”，⁷²而德克萨斯农工大学的研究人员也在开发一个软件平台，旨在自主搜索以前未知的物质。73^这两个项目的部分资金均由美国国防部提供，该部门对任何出现的创新都特别热切。正如亚马逊和谷歌提供的基于云的深度学习工具正在使机器学习在许多商业应用中的部署变得民主化一样，这些工具也将在许多专门的科学研究领域发挥同样的作用。这将使接受过化学或材料科学等领域培训的科学家能够部署人工智能的力量，而无需先成为机器学习专家。换句话说，人工智能正在演变成一种可用工具，可以以更具创造性和针对性的方式运用。

Such initiatives typically require researchers to have a strong technical background in artificial intelligence, but teams at other universities are developing more accessible AI-based tools that are poised to jump-start the discovery of new chemical compounds. Researchers at Cornell University, for example, are working on a project called SARA—Scientific Autonomous Reasoning Agent—which the team hopes will “dramatically accelerate, by orders of magnitude, the discovery and development of new materials,”⁷² while researchers at Texas A&M are likewise developing a software platform designed to autonomously search for previously unknown substances.⁷³ Both projects are funded in part by the U.S. Department of Defense, an especially eager customer for any innovations that emerge. Just as cloud-based deep learning tools offered by Amazon and Google are democratizing the deployment of machine learning in many business applications, tools like these are poised to do the same for many areas of specialized scientific research. This will make it possible for scientists with training in areas like chemistry or materials science to deploy the power of AI without the need to first become machine learning experts. Artificial intelligence, in other words, is evolving into an accessible utility that can be wielded in ever more creative and targeted ways.

一个更加雄心勃勃的方法是将用于发现化学物质的人工智能软件与可以进行物理实验室实验的机器人结合起来。一家朝着这个方向努力的小公司是位于马萨诸塞州剑桥的 Kebotix，这是一家从哈佛大学一家领先的材料科学实验室分离出来的初创公司，它开发了所谓的“世界上第一个用于材料发现的自动驾驶实验室”。该公司的机器人可以自主进行实验，操纵移液器等实验室设备来转移和混合液体，并使用进行化学分析的机器。然后，实验结果由人工智能算法进行分析，算法进而预测最佳行动方案，然后启动更多实验。结果是一个迭代的、自我改进的过程，该公司声称这大大加速了有用新分子的发现^。74

An even more ambitious approach involves integrating AI-based software geared toward the discovery of chemicals with robots that can perform physical laboratory experiments. One small company pushing in this direction is Cambridge, Massachusetts–based Kebotix, a startup that spun out of a leading materials science laboratory at Harvard, which has developed what it calls the “world’s first self-driving lab for materials discovery.” The company’s robots can perform experiments autonomously, manipulating laboratory equipment like pipettes to transfer and combine liquids and accessing machines that perform chemical analysis. Experimental results are then analyzed by artificial intelligence algorithms, which in turn make predictions about the best course of action and then initiate more experiments. The result is an iterative, self-improving process that the company claims dramatically accelerates the discovery of useful new molecules.⁷⁴

在化学与人工智能交叉的领域，许多最令人兴奋且资金最雄厚的机会都在新药的发现和开发中。据一项统计，截至 2020 年 4 月，至少有 230 家初创公司专注于使用人工智能寻找新药。75^斯坦福大学教授、在线教育公司 Coursera 联合创始人达芙妮·科勒 (Daphne Koller) 是将机器学习应用于生物学和生物化学的全球顶尖专家之一。科勒还是硅谷初创公司 insitro 的创始人兼首席执行官，该公司成立于 2018 年，已筹集超过 1 亿美元用于利用机器学习开发新药。困扰整个美国经济的技术创新普遍放缓，在制药行业尤其明显。科勒告诉我：

Many of the most exciting and heavily funded opportunities in the space where chemistry intersects with artificial intelligence are in the discovery and development of new drugs. By one account, as of April 2020, there were at least 230 startup companies focused on using AI to find new pharmaceuticals.⁷⁵ Daphne Koller, a professor at Stanford and the co-founder of the online education company Coursera, is one of the world’s top experts on applying machine learning to biology and biochemistry. Koller is also the founder and CEO of insitro, a Silicon Valley startup, founded in 2018, that has raised over $100 million to pursue new medicines using machine learning. The broad-based slowdown in technological innovation that plagues the American economy as a whole is especially evident in the pharmaceutical industry. Koller told me that:

问题在于，开发新药变得越来越困难：临床试验成功率约为中个位数；开发一种新药的税前研发成本（包括失败）估计超过 25 亿美元。药物开发投资的回报率逐年呈线性下降，一些分析估计，在 2020 年之前将降至零。对此的一个解释是，药物开发现在本质上变得更加困难：许多（也许是大多数）“唾手可得的果实”——换句话说，对大量人群有重大影响的可用药物靶点——已经被发现。如果是这样，那么药物开发的下一阶段将需要专注于更专业化的药物——其效果可能因情况而异，并且仅适用于一小部分患者^。76

The problem is that it is becoming consistently more challenging to develop new drugs: clinical trial success rates are around the mid-single-digit range; the pre-tax R&D cost to develop a new drug (once failures are incorporated) is estimated to be greater than $2.5 [billion]. The rate of return on drug development investment has been decreasing linearly year by year, and some analyses estimate that it will hit zero before 2020. One explanation for this is that drug development is now intrinsically harder: Many (perhaps most) of the “low-hanging fruit”—in other words, druggable targets that have a significant effect on a large population—have been discovered. If so, then the next phase of drug development will need to focus on drugs that are more specialized—whose effects may be context-specific, and which apply only to a subset of patients.⁷⁶

insitro 及其竞争对手的愿景是利用人工智能快速分离有希望的候选药物，并大幅降低开发成本。Koller 说，药物发现是“一条漫长的旅程，路上有多个岔路”，“99% 的路都会让你走进死胡同。”如果人工智能可以提供“一个相对准确的指南针，想想这会对这个过程的成功概率产生什么影响。” ⁷⁷

The vision for insitro and its competitors is to use artificial intelligence to rapidly isolate promising drug candidates and dramatically cut development costs. Drug discovery, Koller says, is “a long journey where you have multiple forks in the road” and “ninety-nine percent of the paths are going to get you to a dead end.” If artificial intelligence can provide “a somewhat accurate compass, think about what that would do to the probability of success of the process.”⁷⁷

此类方法已产生效益。2020 年 2 月，麻省理工学院的研究人员宣布，他们利用深度学习发现了一种强效新抗生素。研究人员建立的人工智能系统能够在几天内筛选出一亿多种潜在化合物。这种新抗生素被科学家命名为 halicin，以电影《2001：太空漫游》中的人工智能系统 HAL 命名。事实证明，这种抗生素对几乎所有类型的细菌都具有致命性，包括对现有药物具有耐药性的菌株。78^这至关重要，因为医学界一直在警告，随着细菌适应现有药物，耐药性细菌（如已经困扰许多医院的“超级细菌”）的危机迫在眉睫。由于开发成本高，利润相对较低，很少有新抗生素正在开发中。即使是那些通过严格而昂贵的测试和监管审批程序的新药，也往往是现有抗生素的变种。相比之下，Halicin 似乎以一种全新的方式攻击细菌，实验表明，这种机制可能对突变特别有抵抗力，而突变通常会随着时间的推移使抗生素的效力降低。换句话说，人工智能已经基于对有意义的创新至关重要的“跳出框框”的探索，提出了一种解决方案。

Approaches like this are already paying dividends. In February 2020, researchers at MIT announced that they had discovered a powerful new antibiotic using deep learning. The AI system built by the researchers was able to sift through more than one hundred million prospective chemical compounds within days. The new antibiotic—which the scientists named halicin after HAL, the artificial intelligence system from 2001: A Space Odyssey—proved lethal to nearly every type of bacteria it was tested against, including strains that are resistant to existing drugs.⁷⁸ This is critical because the medical community has been warning of a looming crisis of drug-resistant bacteria—such as the “superbugs” that already plague many hospitals—as the organisms adapt to existing medications. Because development costs are high and profits relatively low, few new antibiotics are in the development pipeline. Even those new drugs that have made it through the rigorous and expensive testing and regulatory approval process tend to be variations on existing antibiotics. Halicin, in contrast, seems to attack bacteria in a completely novel way, and experiments suggest that the mechanism may be especially resilient to the mutations that generally make antibiotics less effective over time. In other words, artificial intelligence has produced a solution based on the kind of “outside the box” exploration that is critical to meaningful innovation.

另一个重要里程碑也于 2020 年初宣布，来自英国初创公司 Exscientia，该公司利用机器学习发现了一种治疗强迫症的新药。该公司表示，该项目的初始开发仅用了一年时间——大约是传统技术所需时间的五分之一——并声称这是第一个进入临床试验的人工智能发现的药物^。79

Another important milestone, also announced in early 2020, came from the U.K.-based startup company Exscientia, which used machine learning to discover a new drug for treating obsessive compulsive disorder. The company says the project’s initial development took just one year—about one fifth the time that would be typical for traditional techniques—and claims it is the first AI-discovered drug to enter clinical trials.⁷⁹

正如我们在第 1 章中看到的那样，人工智能在生化研究中应用的一个特别值得注意的成就是 DeepMind 于 2020 年 11 月宣布的蛋白质折叠突破。DeepMind 并没有试图发现一种特定的药物，而是将其技术用于在更基本的层面上获得理解。2018 年末，DeepMind 将其 AlphaFold 系统的早期版本送入了两年一度的全球竞赛，即结构预测关键评估 (CASP)。来自世界各地的团队使用基于计算和人类直觉的各种技术来尝试预测蛋白质的折叠方式。AlphaFold 以较大优势赢得了 2018 年的比赛，但即使占了上风，它也只能对 43 个蛋白质序列中的 25 个做出正确的预测。换句话说，AlphaFold 的这个初步版本还不够准确，还不能成为真正有用的研究工具。⁸⁰事实上，DeepMind 能够不断完善其技术，以至于仅在两年后，许多科学家就宣布蛋白质折叠问题已“解决”，我认为，这一事实特别生动地表明了人工智能的具体应用可能会继续快速发展。

As we saw in Chapter 1, an especially notable achievement in the application of artificial intelligence to biochemical research was DeepMind’s protein folding breakthrough announced in November 2020. Rather than attempting to discover a specific drug, DeepMind has instead deployed its technology to gain understanding at a more fundamental level. In late 2018, DeepMind entered an earlier version of its AlphaFold system in a biennial global contest known as the Critical Assessment of Structure Prediction, or CASP. Teams from around the world used a variety of techniques based on both computation and human intuition to attempt to predict the way proteins fold. AlphaFold won the 2018 contest by a wide margin, but even while prevailing, it was able to make the best prediction for only twenty-five of the forty-three protein sequences correctly. In other words, this preliminary version of AlphaFold was not yet accurate enough to be a truly useful research tool.⁸⁰ The fact that DeepMind was able to refine its technology to the point where a number of scientists declared the protein folding problem to be “solved” just two years later is, I think, an especially vivid indication of just how rapidly specific applications of artificial intelligence are likely to continue advancing.

除了使用机器学习来发现新药和其他化学化合物外，人工智能在科学研究中最有前景的普遍应用可能是吸收和理解不断激增的已发表研究成果。仅在 2018 年，就有超过 300 万篇科学论文发表在 40,000 多种不同的期刊上。81理解如此规模的信息远远超出了任何个人思维的能力，因此人工智能可以说^是我们掌握的唯一能够实现某种整体理解的工具。

Aside from using machine learning to discover new drugs and other chemical compounds, the most promising general application of artificial intelligence to scientific research may be in the assimilation and understanding of the continuously exploding volume of published research. In 2018 alone, more than three million scientific papers were published in more than 40,000 separate journals.⁸¹ Making sense of information on that scale is so far beyond the capability of any individual human mind that artificial intelligence is arguably the only tool at our disposal that could lead to some sort of holistic comprehension.

基于深度学习最新进展的自然语言处理系统正在被部署用于提取信息、识别研究中不明显的模式，并通常建立可能仍然模糊的概念联系。IBM 的 Watson 技术继续成为这一领域的重要参与者。另一个项目 Semantic Scholar 由位于西雅图的艾伦人工智能研究所于 2015 年发起。Semantic Scholar 提供人工智能搜索和信息提取服务，涵盖几乎所有科学研究领域的 1.86 亿多篇已发表的研究论文^。82

Natural language processing systems based on the latest advances in deep learning are being deployed to extract information, identify non-obvious patterns across research studies and generally make conceptual connections that might otherwise remain obscure. IBM’s Watson technology continues to be one important player in this space. Another project, Semantic Scholar, was initiated by the Seattle-based Allen Institute for Artificial Intelligence in 2015. Semantic Scholar offers AI-enabled search and information extraction across more than 186 million published research papers in virtually every scientific field of study.⁸²

2020 年 3 月，艾伦研究所与微软、美国国家医学图书馆、白宫科技办公室、亚马逊 AWS 部门等其他组织联合创建了 COVID-19 开放研究数据集，这是一个可搜索的与冠状病毒大流行相关的科学论文数据库。83^该技术使科学家和医疗保健提供者能够快速获取广泛科学领域特定问题的答案，包括病毒的生物化学、流行病学模型和疾病的治疗。截至 2021 年 4 月，该数据库包含超过 280,000 篇科学论文，并被科学家和医生大量使用^。84

In March 2020, the Allen Institute joined with a consortium of other organizations including Microsoft, the National Library of Medicine, the White House Office of Science and Technology, Amazon’s AWS division and others to create the COVID-19 Open Research Dataset, a searchable database of scientific papers relating to the coronavirus pandemic.⁸³ The technology enables scientists and healthcare providers to rapidly access answers to specific questions in a broad range of scientific areas, including the biochemistry of the virus, epidemiological models, and treatment of the disease. As of April 2021, the database contained more than 280,000 scientific papers and was being heavily used by scientists and doctors.⁸⁴

此类举措具有巨大潜力，有望成为加速新创意产生的关键工具。然而，这项技术仍处于起步阶段，真正的进步可能需要克服通往更通用的机器智能道路上的至少一些障碍，我们将在第5 章深入探讨这一主题。不难想象，一个真正强大的系统可以充当科学家的智能研究助手，提供进行真诚对话、发挥创意和积极提出新探索途径的能力。

Initiatives like these have enormous potential to be crucial tools in accelerating the generation of new ideas. The technology remains in its infancy, however, and real progress will likely require surmounting at least some of the hurdles on the path to more general machine intelligence, a subject we’ll delve into in Chapter 5. It’s easy to imagine that a truly powerful system could step into the role of an intelligent research assistant for scientists, offering the ability to engage in genuine conversation, play with ideas and actively suggest new avenues for exploration.

不过，我认为重要的是对最终可能发生的事情保持慎重和现实的看法。所有这些都不意味着人工智能将成为加速创新的灵丹妙药，也不意味着我们应该期望在加速的时间内持续取得成果。毕竟，科学从根本上讲是关于实验的，而进行和评估实验结果需要时间。在某些情况下，科学方法确实可以加速，也许是通过使用实验室机器人，甚至通过在模拟环境中高速进行一些实验。

Still, I think it’s important to maintain a measured and realistic view of what might eventually be possible. None of this implies that artificial intelligence will be a panacea for turbocharging innovation or that we should expect results to consistently be achieved on an accelerated time frame. Science is, after all, fundamentally about experimentation, and conducting and evaluating the outcomes of experiments takes time. In some cases, the scientific method can indeed be accelerated, perhaps through the use of laboratory robots or even by performing some experiments at high speed in simulated environments.

然而，在医学和生物学等领域，实验通常必须在活体生物体内进行，而在这里，大幅加快这一过程的潜力相当有限。COVID-19 疫苗的成功研发使这一现实成为人们关注的焦点。科学家们在获得病毒遗传密码后的数周内就能够配制出候选疫苗。等待可用疫苗的时间太长几乎完全是因为需要在动物和人类身上进行大量测试，以及需要提高生产能力以生产所需的数十亿剂疫苗。事实是，即使我们能够使用真正先进的科幻级人工智能，也完全不清楚这项技术是否能够在更短的时间内提供疫苗。这是我对库兹韦尔的说法持怀疑态度的原因之一，库兹韦尔声称人工智能将很快导致人类寿命大幅延长。即使人工智能确实有助于在这个领域产生强大的新想法，我们如何在不等待多年甚至几十年才能得到确凿结果的情况下测试任何由此产生的治疗方法的安全性和有效性？可以肯定的是，监管改革有很多机会可以简化新药和新疗法的审批，但最终，即使是最聪明、最有创造力的科学家也必须等待实验结果来证实其想法的真实性。

In fields like medicine and biology, however, experiments often must be conducted within living organisms, and here, the potential for dramatically speeding up the process is quite limited. The successful quest for COVID-19 vaccines brings this reality into sharp focus. Scientists were able to formulate vaccine candidates within weeks of obtaining the virus’s genetic code. The long wait for serviceable vaccines was almost entirely due to the need for extensive testing in both animals and humans, along with the need to ramp up manufacturing capacity to produce the billions of doses that are required. The truth is that even if we had access to truly advanced, science fiction–level artificial intelligence, it is not at all clear that the technology could have delivered a vaccine in a dramatically shorter time frame. This is one of the reasons that I’m skeptical of Kurzweillian claims that artificial intelligence will soon lead to a dramatic lengthening of the human lifespan. Even if AI does help generate powerful new ideas in this space, how will we test any resulting treatments for both safety and efficacy without waiting many years or even decades for conclusive results? To be sure, there are many opportunities for regulatory reform that might streamline the approval of new drugs and treatments, but at the end of the day, even the most intelligent and creative scientists must wait for the experimental results that confirm the veracity of their ideas.

本章旨在简要介绍人工智能的一些最有趣和最重要的应用，同时强调人工智能在短期内可能带来颠覆的领域以及我们可能需要等待更长时间的领域。这份清单远非详尽无遗。人工智能最终将触及并改变几乎所有事物。

THE INTENT OF this chapter has been to offer a brief tour of some of the most interesting and consequential applications of artificial intelligence while highlighting areas where AI seems likely to be disruptive in the near term and where we may have a longer wait. The list is nowhere close to being exhaustive. Artificial intelligence will eventually touch and transform virtually everything.

人工智能正在迅速发展成为一种类似电力的公用事业，这一论点有效地体现了该技术的潜在影响力和变革性。然而，与电力相比，人工智能是一项复杂得多、动态性更强的技术，它将不断改进，同时提供几乎无限数量的不断变化的功能。为了了解这种新公用事业的真正潜力，我们需要深入研究人工智能的科学和历史，看看这个领域是如何发展的，未来面临的挑战，以及在技术不断进步的过程中将塑造它的竞争理念。所有这些都将是接下来两章的主题。

The argument that artificial intelligence is rapidly evolving into an electricity-like utility effectively captures the potential reach and transformative nature of the technology. Compared to electricity, however, AI is a vastly more complex and dynamic technology that will continuously improve while delivering a nearly limitless number of ever-changing capabilities. In order to understand the true potential of this new utility, we need to delve into the science and history of artificial intelligence and see how the field is evolving, the challenges that lie ahead and the competing ideas that will shape the technology as it continues to progress. All this will be the subject of the next two chapters.

脚注

Footnote

ⁱ 2020 年 10 月，特斯拉确实发布了所谓的“全自动驾驶套件”的早期版本。该软件已通过下载提供给有限数量的特斯拉车主，并计划在接下来的几个月内扩大可用性。该软件提供自动停车和有限的城市街道导航能力等功能，但目前还远远达不到可以合理称为“全自动驾驶”的水平。特斯拉承诺升级该套件，并宣布未来将涨价，以激励车主购买早期版本。美国国家公路交通安全管理局注意到了这一点，并宣布将“密切关注新技术”，并“将毫不犹豫地采取行动保护公众免受不合理的安全风险”。（见第 3 章尾注 2。）

ⁱ In October 2020, Tesla did, in fact, release an early version of what it calls a “Full Self-Driving Package.” The software was made available to a limited number of Tesla owners via download, with plans to expand availability over the ensuing months. The software provides features such as automatic parking and a limited ability to navigate city streets but is currently nothing close to what could reasonably be called “full self-driving.” Tesla has promised to upgrade the package and has announced future price increases in order to incentivize owners to purchase an early version. The National Highway Traffic Safety Administration took note and declared that it would “monitor the new technology closely” and that it “will not hesitate to take action to protect the public against unreasonable risks to safety.” (See endnote 2, Chapter 3.)

⁸⁴

第四章

CHAPTER 4 打造智能机器的探索

THE QUEST TO BUILD INTELLIGENT MACHINES

图灵奖被公认为计算机界的“诺贝尔奖”。图灵奖以传奇数学家和计算机科学家艾伦·图灵的名字命名，每年由计算机协会颁发，代表了那些致力于推动该领域发展的人的最高成就。与诺贝尔奖一样，图灵奖的奖金为 100 万美元，主要由谷歌资助。

THE A.M. TURING AWARD IS GENERALLY RECOGNIZED AS THE “Nobel Prize” of computing. Named after the legendary mathematician and computer scientist Alan Turing and awarded annually by the Association for Computing Machinery, the Turing Award represents the pinnacle of achievement for those who have devoted their careers to advancing the state of the field. Like the Nobel, the Turing prize comes with a $1 million financial award, which is funded primarily by Google.

2019 年 6 月，2018 年图灵奖颁给了 Geoffrey Hinton、Yann LeCun 和 Yoshua Bengio 三人，以表彰他们为深度神经网络的发展做出的终身贡献。这项技术（也称为深度学习）在过去十年中彻底改变了人工智能领域，并取得了不久前还被认为是科幻小说的进步。

In June 2019, the 2018 Turing Award was awarded to three men—Geoffrey Hinton, Yann LeCun and Yoshua Bengio—in recognition of their lifetime contributions to the advancement of deep neural networks. This technology—also known as deep learning—has, over the past decade, revolutionized the field of artificial intelligence and produced advances that just a short time ago would have been considered science fiction.

特斯拉司机经常让他们的汽车在高速公路上自动驾驶。谷歌翻译可以即时生成可用的文本，即使是我们很少听说过的生僻语言也可以，微软等公司也展示了实时机器翻译，可以将口语中的中文翻译成英文。孩子们成长在一个与亚马逊 Alexa 交谈习以为常的世界里，父母们担心这些互动是否健康。所有这些进步——以及许多其他进步——都是由深度神经网络驱动的。

Tesla drivers routinely let their cars navigate highways autonomously. Google Translate instantly produces usable text, even in obscure languages that few of us have heard of, and companies like Microsoft have demonstrated real-time machine translation that renders spoken Chinese into English. Children are growing up in a world where it is routine to converse with Amazon’s Alexa, and parents are worrying about whether these interactions are healthy. All of these advances—and a multitude of others—are powered by deep neural networks.

深度学习的基本思想已经存在了几十年。20 世纪 50 年代末，康奈尔大学心理学家弗兰克·罗森布拉特 (Frank Rosenblatt) 构想出了“感知器”，这是一种电子设备，其工作原理与大脑中的生物神经元类似。罗森布拉特表明，由感知器组成的简单网络可以训练来执行基本的模式识别任务，例如解读数字图像。

The basic idea underlying deep learning has been around for decades. In the late 1950s, Frank Rosenblatt, a psychologist at Cornell University, conceived the “perceptron,” an electronic device that operated on principles similar to those of biological neurons in the brain. Rosenblatt showed that simple networks made up of perceptrons could be trained to perform basic pattern recognition tasks, such as deciphering images of numerical digits.

罗森布拉特最初在神经网络方面的工作引起了人们的热情，但由于未能取得重大进展，该技术最终被其他方法所取代。只有一小部分研究人员，尤其是 2018 年图灵奖的三位获奖者，继续专注于神经网络。在计算机科学家中，这项技术被视为研究的死水和可能的职业死胡同。

Rosenblatt’s initial work on neural networks generated enthusiasm, but as significant progress failed to materialize, the technique was eventually pushed aside by other approaches. Only a small group of researchers, including especially the three winners of the 2018 Turing Award, continued to focus on neural networks. Among computer scientists, the technology came to be viewed as a research backwater and a likely career dead end.

2012 年，多伦多大学 Geoff Hinton 研究实验室的一支团队参加了 ImageNet 大规模视觉识别挑战赛，一切都发生了改变。在这项年度赛事中，来自世界许多一流大学和公司的团队竞相设计一种算法，该算法可以正确标记从海量照片数据库中选出的图像。其他团队使用传统的计算机编程技术，而 Hinton 的团队则推出了一个“深度”（或多层）神经网络，该网络已在数千张示例图像上进行过训练。多伦多大学团队在比赛中大获全胜，全世界开始意识到深度学习的力量。

Everything changed in 2012 when a team from Geoff Hinton’s research lab at the University of Toronto entered the ImageNet Large Scale Visual Recognition Challenge. In this annual event, teams from many of the world’s leading universities and corporations competed to design an algorithm that could correctly label images selected from a massive database of photographs. While other teams used traditional computer programming techniques, Hinton’s team unleashed a “deep” (or many-layered) neural network that had been trained on thousands of example images. The University of Toronto team blew the doors off the competition, and the world woke up to the power of deep learning.

此后的几年里，几乎每一家大型科技公司都对深度学习进行了大规模投资。谷歌、Facebook、亚马逊和微软，以及中国科技领军企业百度、腾讯和阿里巴巴，都将深度神经网络作为其产品、运营和商业模式的核心。计算机硬件行业也在发生转变，英伟达和英特尔等公司竞相开发优化神经网络性能的计算机芯片。由于各公司都在争夺有限的人才，深度学习专家的薪酬高达七位数，待遇堪比明星运动员。

In the years since, virtually every major technology company has made massive investments in deep learning. Google, Facebook, Amazon and Microsoft, as well as the Chinese tech leaders Baidu, Tencent and Alibaba, have made deep neural networks absolutely central to their products, operations and business models. The computer hardware industry is also being transformed, with companies like NVIDIA and Intel competing to develop computer chips that optimize the performance of neural networks. Experts in deep learning command seven-figure compensation packages and are treated like star athletes as companies compete for a limited pool of talent.

尽管过去十年人工智能的进步既非凡又史无前例，但这一进步在很大程度上是由越来越大的数据库规模推动的，这些数据库被运行在越来越快的计算机硬件上的神经学习算法所吞噬。人工智能专家们越来越意识到，这种方法是不可持续的，需要将全新的想法注入技术中，才能继续推动技术的发展。在深入探讨人工智能的未来之前，让我们先简单回顾一下这一切是如何开始的，这个领域迄今为止走过的道路，以及在过去几年中取得如此革命性进展的深度学习系统是如何工作的。正如我们所看到的，自人工智能研究的早期开始，就以两种完全不同的智能机器构建方法之间的竞争为标志。这两种思想流派之间的紧张关系再次成为焦点，很可能将在未来几年和几十年内影响该领域的发展方式。

Though advances in artificial intelligence over the past decade have been both extraordinary and unprecedented, this progress has largely been driven by scaling up to ever larger troves of data being gobbled up by neural learning algorithms running on faster and faster computer hardware. There’s a growing sense among AI experts that this approach is not sustainable and that entirely new ideas will need to be injected into the technology in order to continue carrying things forward. Before delving into the possible future of AI, let’s briefly look at how it all began, the path that the field has traced up until now and how the deep learning systems that have produced such revolutionary progress over the past few years actually work. As we’ll see, since its earliest days, research into artificial intelligence has been marked by competition between two entirely different approaches to building smart machines. Tension between these two schools of thought is once again coming to the forefront and may well be poised to shape the way the field progresses in the coming years and decades.

机器可以思考吗？

CAN MACHINES THINK?

早在第一台电子计算机发明之前，人们就想象过拥有像人类一样思考和行动能力的机器。1863 年，英国作家塞缪尔·巴特勒给新西兰基督城当地报纸的编辑写了一封信。这封信的标题是《机器中的达尔文》，信中设想了“活机器”有朝一日可能会进化到与人类相匹敌甚至取代人类。巴特勒呼吁立即对这种新兴的机械物种发动战争，宣称“每种机器都应该被摧毁”，¹考虑到 1863 年的信息技术状况，这种担忧似乎有点为时过早，但它勾勒出了一个一再重复的故事，最近一次出现在《终结者》和《黑客帝国》等电影中。巴特勒的担忧并不局限于科幻小说。人工智能的最新进展已使埃隆·马斯克和已故的斯蒂芬·霍金等知名人物发出警告，称与 150 多年前巴特勒担心的情况非常相似的情景将会出现。

Machines with the ability to think and act like humans inhabited the imagination long before the invention of the first electronic computers. In 1863, the English author Samuel Butler wrote a letter to the editor of the local newspaper in Christchurch, New Zealand. Entitled “Darwin Among the Machines,” the letter envisioned “living machinery” that might someday evolve to match and perhaps even supplant human beings. Butler called for an immediate war against this emerging mechanical species, proclaiming that “every machine of every sort should be destroyed,”¹ a concern that seems a bit premature given the state of information technology in 1863, but which sketched out a narrative that has been repeated again and again, most recently in movies like The Terminator and The Matrix. Nor are Butler’s fears confined to science fiction. Recent advances in AI have led prominent figures like Elon Musk and the late Stephen Hawking to warn of scenarios remarkably similar to what Butler worried about more than 150 years ago.

关于人工智能成为一门严肃学科的确切时间，人们的看法不一。我认为人工智能的起源可以追溯到 1950 年。那一年，才华横溢的数学家艾伦·图灵发表了一篇题为“计算机器和智能”的科学论文，提出了“机器能思考吗？”的问题。^图灵在论文中发明了一种测试，该测试基于聚会上流行的游戏，至今仍是判断机器是否真正智能的最常用方法。图灵于 1912 年出生于伦敦，在计算理论和算法性质方面做出了开创性的工作，被公认为计算机科学之父。图灵最重要的成就是在 1936 年，即他从剑桥大学毕业仅两年后，他提出了如今所谓的“通用图灵机”的数学原理——这基本上是所有现实世界中计算机的概念蓝图。图灵在计算机时代伊始就清楚地认识到，机器智能是电子计算的合乎逻辑且可能不可避免的延伸。

Opinions differ as to exactly when artificial intelligence became a serious field of study. I would mark the origin as 1950. In that year, the brilliant mathematician Alan Turing published a scientific paper entitled “Computing Machinery and Intelligence” that asked the question “Can machines think?”² In his paper Turing invented a test, based on a game that was popular at parties, which is still the most commonly cited method for determining if a machine can be considered to be genuinely intelligent. Turing, born in London in 1912, did groundbreaking work on the theory of computation and the nature of algorithms, and is generally regarded as the founding father of computer science. Turing’s most important accomplishment came in 1936, just two years after he graduated from the University of Cambridge, when he laid out the mathematical principles for what is today called a “universal Turing machine”—essentially the conceptual blueprint for every real-world computer that has ever been built. Turing clearly understood at the very inception of the computer age that machine intelligence was a logical and perhaps inevitable extension of electronic computation.

“人工智能”一词由当时还是达特茅斯学院年轻数学教授的约翰·麦卡锡首创。1956 年夏天，麦卡锡帮助安排了达特茅斯学院新罕布什尔校区的人工智能夏季研究项目。这是一个为期两个月的会议，邀请了新兴领域的领军人物参加。目标既雄心勃勃又乐观；会议提案宣称“将尝试找到如何让机器使用语言、形成抽象和概念、解决目前只有人类才能解决的各种问题并自我改进”，并承诺组织者相信“如果一群精心挑选的科学家在一个夏天内共同努力，就可以在这些问题中的一个或多个方面取得重大进展”。3与会者包括马^文·明斯基，他与麦卡锡一起成为世界上最著名的人工智能研究人员之一，并在麻省理工学院建立了计算机科学和人工智能实验室，还有克劳德·香农，一位传奇的电气工程师，他制定了电子通信的基础信息理论原理并使互联网成为可能。

The phrase “artificial intelligence” was coined by John McCarthy, who was then a young mathematics professor at Dartmouth College. In the summer of 1956, McCarthy helped arrange the Dartmouth Summer Research Project on Artificial Intelligence at the college’s New Hampshire campus. This was a two-month conference to which the leading lights of the newly emerging field were invited. The goals were both ambitious and optimistic; the conference proposal declared that “an attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves” and promised that the organizers’ believed a “significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.”³ Attendees included Marvin Minsky, who along with McCarthy became one of the world’s most celebrated AI researchers and founded the Computer Science and Artificial Intelligence Lab at MIT, and Claude Shannon, a legendary electrical engineer who formulated the principles of information theory that underlie electronic communication and make the internet possible.

然而，这位最聪明的人却没有出席达特茅斯会议。两年前，艾伦·图灵自杀了。图灵因同性恋关系被控违反当时英国的“猥亵”法，他被要求在监禁或通过强制注射雌激素进行化学阉割之间做出选择。在选择了第二个选项后，他因抑郁而于 1954 年自杀。对于计算机科学和人工智能等新兴领域而言，这种损失将是无法估量的。图灵去世时只有 41 岁。在一个更加公正的世界里，他几乎肯定会活着看到个人电脑的出现，很可能看到互联网的兴起，甚至可能看到随后的许多创新。没有人能说出图灵在那几十年里做出了什么贡献，或者现在人工智能领域可能取得了多大的进展，但对于该领域和全人类而言，智力损失可能是惊人的。

The brightest mind, however, was notably absent from the Dartmouth conference. Alan Turing had committed suicide two years earlier. Prosecuted for a same-sex relationship under the “indecency” laws then in force in Britain, Turing was given a choice between imprisonment or chemical castration through the forced introduction of estrogen. Depressed after selecting the second option, he took his own life in 1954. For the emergent fields of computer science and artificial intelligence, the loss would be incalculable. Turing was only forty-one when he died. In a more just world, he would have almost certainly lived to see the advent of the personal computer and quite possibly the rise of the internet, and perhaps even many of the innovations that followed. No one can say what contributions Turing might have made over those decades, or how much further along the field of artificial intelligence might now be, but the intellectual loss to the field, and to all of humanity, is likely staggering.

达特茅斯会议之后的几年里，人工智能领域取得了快速发展。计算机变得越来越强大，取得了重大突破，开发出了可以解决越来越多问题的算法。人工智能作为一个研究领域被引入美国各大学，并建立了许多人工智能研究实验室。

The field of artificial intelligence made rapid progress in the years following the Dartmouth conference. Computers were becoming more capable, important breakthroughs were made and algorithms were developed that could solve an increasing range of problems. Artificial intelligence as a field of study was introduced at universities across the United States, and a number of AI research labs were established.

推动这一进步的最重要因素之一是美国政府，尤其是五角大楼的巨额投资。这些投资大部分来自美国高级研究计划局（ARPA）。APRA 资助的一个特别重要的研究中心是斯坦福研究所，该研究所后来从斯坦福大学分离出来，成为 SRI 国际研究所。SRI 的人工智能中心成立于 1966 年，在语言翻译和语音识别等领域做出了开创性的工作。该实验室还创造了第一个真正自主的机器人，这种机器能够将人工智能推理转化为与环境的物理交互。成立近半个世纪后，SRI 的人工智能中心将成立一家初创公司，推出一款名为 Siri 的新个人助理，该公司于 2010 年被苹果收购。

One of the most important enablers of this progress was massive investment from the U.S. government, especially the Pentagon. Much of this was funneled through the Advanced Research Projects Agency, or ARPA. One especially important center of APRA-funded research was the Stanford Research Institute, which later separated from Stanford University to become SRI International. SRI’s Artificial Intelligence Center, established in 1966, did groundbreaking work in areas like language translation and speech recognition. The lab also created the first truly autonomous robot, a machine capable of turning AI-powered reasoning into physical interaction with the environment. Nearly half a century after its founding, SRI’s Artificial Intelligence Center would spin off a startup company with a new personal assistant called Siri that would be acquired by Apple in 2010.

然而，进步很快就导致了过度兴奋、过大的承诺和不切实际的期望。1970 年，《生活》杂志发表了一篇关于 SRI 开发的机器人的文章，称其为世界上第一个“电子人”。当时是麻省理工学院明星人工智能研究员的马文·明斯基向文章作者布拉德·达拉赫“平静地确信”道：

Progress, however, soon led to overexuberance, outsized promises and unrealistic expectations. In 1970, LIFE published an article about the robot developed at SRI, calling it the world’s first “electronic person.” Marvin Minsky, who was then a star AI researcher at MIT, declared to the article’s author, Brad Darrach, with “quiet certitude”:

3 到 8 年后，我们将拥有一台拥有普通人类一般智力的机器。我的意思是，这台机器将能够阅读莎士比亚、给汽车加油、玩办公室政治、讲笑话、打架。到那时，机器将开始以惊人的速度自我教育。几个月后，它将达到天才的水平，再过几个月，它的能力将不可估量^。4

In from three to eight years we will have a machine with the general intelligence of an average human being. I mean a machine that will be able to read Shakespeare, grease a car, play office politics, tell a joke, have a fight. At that point the machine will begin to educate itself with fantastic speed. In a few months it will be at genius level and a few months after that its powers will be incalculable.⁴

Darrach 与其他 AI 研究人员核实了这一说法，并被告知明斯基的 3 到 8 年时间框架可能有点乐观。他们说，这可能需要 15 年，但“所有人都同意会有这样一台机器，它可以促成第三次工业革命，消灭战争和贫困，并延续科学、教育和艺术的数百年发展。” ⁵

Darrach checked this statement with other AI researchers and was told that perhaps Minsky’s three-to-eight-year time frame was a bit optimistic. It might take fifteen years, they said, but “all agreed that there would be such a machine and that it could precipitate the Third Industrial Revolution, wipe out war and poverty, and roll up centuries of growth in science, education and the arts.”⁵

随着人们发现这些预测大错特错，而且开发能够执行远没有那么宏伟的任务的人工智能系统也比预期的要困难得多，该领域的热情开始消退。到 1974 年，投资者的幻灭，尤其是那些在资金方面发挥了巨大作用的政府机构，给该领域蒙上了一层阴影，也给许多人工智能研究人员的职业前景蒙上了一层阴影。纵观人工智能的发展历史，该领域一直遭受着一种集体躁郁症的困扰，在高度繁荣和快速发展的时期，有时会出现长达数十年的幻灭和低投资时期，这被称为“人工智能寒冬”。

As it became evident that such predictions were wildly off the mark, and that building AI systems capable of performing even far less ambitious tasks was proving to be much more difficult than expected, enthusiasm began to drain away from the field. By 1974, disillusionment among investors, especially the government agencies that played an outsized funding role, cast a pall over the field—and over the career prospects of many AI researchers. Throughout its history, the field of artificial intelligence has suffered from a kind of collective bipolar disorder, with periods of high exuberance and rapid progress punctuated by sometimes decades-long stretches of disillusionment and low investment that have come to be called “AI winters.”

该领域周期性地陷入人工智能寒冬，部分原因可能是人们没有意识到人工智能想要解决的问题到底有多难。然而，另一个关键因素是人们没有意识到 20 世纪 90 年代之前的计算机速度有多慢。在摩尔定律的不懈推动下，人们花了几十年的时间才研发出硬件，让 1956 年达特茅斯会议参与者的梦想得以实现。

The field’s periodic plunge into AI winters likely resulted in part from a lack of appreciation of just how difficult the problems that AI aims to solve really are. Another critical factor, however, was a simple failure to recognize how genuinely slow computers were prior to the 1990s. It would take decades of progress under the relentless regime of Moore’s Law to deliver hardware that would begin to put the dreams of the 1956 Dartmouth conference participants within reach.

速度更快的计算机硬件的出现，在 20 世纪 90 年代末带来了一些重大进展。1997 年 5 月，IBM 的“深蓝”计算机在六局比赛中险胜国际象棋世界冠军加里·卡斯帕罗夫。虽然这被普遍视为人工智能的胜利，但事实上，这主要是通过利用蛮力计算的能力实现的。在“深蓝”冰箱大小的定制硬件上运行的专门算法能够预见未来，快速筛选大量潜在动作，即使是最聪明的人脑也无法做到这一点。

The arrival of that faster computer hardware led to some dramatic advances in the late 1990s. In May 1997, IBM’s Deep Blue computer narrowly defeated the world chess champion, Garry Kasparov, in a six-game match. Though this was generally heralded as a triumph for artificial intelligence, in fact it was primarily a feat accomplished by leveraging the power of brute computation. The specialized algorithms that ran on Deep Blue’s refrigerator-sized, custom-designed hardware were able to look far ahead, rapidly sifting through a multitude of potential moves in a way that would have been impossible for even the most capable human mind.

2011 年，IBM 再次凭借 Watson 的问世而大获全胜。Watson 是一款在电视游戏节目《危险边缘！》（Jeopardy!）中轻松击败世界顶级选手的机器。从很多方面来看，这都是一项令人印象深刻的壮举，因为它需要理解自然语言，甚至包括理解笑话和双关语的能力。与 Deep Blue 不同，Watson 是一个可以超越严格规则的棋盘游戏局限并处理看似无限的信息的系统。Watson 在《危险边缘！》中获胜，因为它同时部署了一组智能算法，这些算法在玩游戏时处理大量数据（通常来自维基百科文章），以确定正确的答案。

IBM triumphed again in 2011 with the advent of Watson, a machine that easily defeated the world’s top contestants on the TV game show Jeopardy! In many ways this was a far more impressive feat because it required an understanding of natural language that included even the ability to navigate jokes and puns. Unlike Deep Blue, Watson was a system that could go beyond the confines of a board game with rigidly defined rules and deal with a seemingly limitless body of information. Watson won at Jeopardy! by simultaneously deploying a swarm of smart algorithms that coursed through reams of data, often drawn from Wikipedia articles, to determine the correct responses as it played the game.

沃森预示着一个新时代的到来，预示着机器最终将开始解析语言并真正与人类互动，但 2011 年也标志着人工智能底层技术发生巨大转变的开始。沃森依靠机器学习算法，该算法使用统计技术来理解信息，但在接下来的几年里，另一种机器学习——直接基于弗兰克·罗森布拉特半个多世纪前构想的感知器——将再次成为焦点，然后迅速崛起并主导人工智能领域。

Watson heralded a new age and portended machines that would finally begin to parse language and truly engage with humans, but 2011 would also mark the beginning of a dramatic shift in the underlying technology of artificial intelligence. Watson relied on machine learning algorithms that used statistical techniques to make sense of information, but over the next few years, another kind of machine learning—based directly on the perceptron conceived by Frank Rosenblatt more than half a century earlier—would once again come to the forefront and then rapidly rise to dominate the field of artificial intelligence.

联结主义与符号主义人工智能以及深度学习的兴起

CONNECTIONIST VS. SYMBOLIC AI AND THE RISE OF DEEP LEARNING

几十年来，人工智能领域经历了兴衰起伏，研究重点在两种普遍的哲学之间摇摆不定，这两种哲学强调构建更智能机器的截然不同的方法。一种思想流派源于罗森布拉特在 20 世纪 50 年代对神经网络的研究。该流派的追随者认为，智能系统应该以大脑的基本结构为模型，并应该利用基于生物神经元的深度连接组件。这种方法后来被称为“联结主义”，强调学习是智能的核心能力，并认为如果机器能够有效地从数据中学习，那么人类大脑所表现出的其他能力最终也可能会显现出来。毕竟，有强有力的证据证明这种模型是有效的：众所周知，人类大脑本身完全由一个难以理解的复杂系统组成，这些系统由相互连接的生物神经元组成。

Even as the general field of artificial intelligence traced its boom-and-bust path over the decades, the research focus swung between two general philosophies that emphasized contrasting approaches to building more intelligent machines. One school of thought grew out of Rosenblatt’s work on neural networks in the 1950s. Its adherents believed that an intelligent system should be modeled on the fundamental architecture of the brain and that it should utilize deeply connected components based loosely on biological neurons. This approach, which came to be called “connectionism,” emphasized learning as the central capability of intelligence and argued that if a machine could be made to efficiently learn from data, then the other capabilities exhibited by the human brain might eventually emerge as well. There was, after all, strong evidence for the effectiveness of this model: the human brain itself was known to be composed entirely of an incomprehensibly complex system of interconnected biological neurons.

竞争阵营的研究人员采用“符号”方法，强调逻辑和推理的应用。对于符号学家来说，学习并不那么重要。相反，智能的关键在于通过推理、决策和行动利用知识的能力。符号学家不是设计能够自我学习的算法，而是手动将信息直接编码到他们构建的系统中，这种做法催生了一个被称为“知识工程”的领域。

The competing camp consisted of researchers who embraced a “symbolic” approach that emphasized the application of logic and reasoning. For the symbolists, learning was not so important. Instead, the key to intelligence was the ability to leverage knowledge through reasoning, decision-making and action. Rather than designing algorithms that could learn by themselves, symbolists manually encoded information directly into the systems they built, a practice that gave rise to a field known as “knowledge engineering.”

符号人工智能是几乎所有早期人工智能实际应用的驱动力。例如，知识工程师与医生合作，能够构建使用决策树算法来诊断疾病的系统。这种医疗专家系统产生了好坏参半的结果，而且往往被证明是不灵活和不可靠的。然而，在许多其他应用中，例如喷气式飞机上使用的自动驾驶系统，通过研究专家系统而开发的技术已逐渐成为软件设计的常规组成部分，不再被称为“人工智能”。

Symbolic AI was the engine that powered nearly all early practical applications of artificial intelligence. Knowledge engineers, working with doctors, for example, were able to build systems that attempted to diagnose illnesses using algorithms that employed decision trees. Such medical expert systems produced mixed results and often proved to be inflexible and unreliable. However, in many other applications, such as the autopilot systems used on jet aircraft, the techniques that were developed through research into expert systems have gradually become routine components of software design and are no longer labeled as “artificial intelligence.”

联结主义起源于旨在了解人类大脑功能的研究。20 世纪 40 年代，沃伦·麦卡洛克和沃尔特·皮茨提出了人工神经网络的概念，将其作为大脑中生物神经元运作方式的一种计算近似。6^受过心理学训练并在康奈尔大学心理学系授课的弗兰克·罗森布拉特后来将这些想法融入了他的感知器中。

Connectionism had its origins in research that aimed to understand the function of the human brain. In the 1940s, Warren McCulloch and Walter Pitts introduced the idea of an artificial neural network as a kind of computational approximation for the way the biological neurons in the brain operated.⁶ Frank Rosenblatt, who was trained as a psychologist and lectured in the psychology department at Cornell, later incorporated these ideas into his perceptron.

感知器能够执行基本的模式识别任务，例如通过连接到设备上的摄像头识别印刷字符。发明者兼作家雷·库兹韦尔 (Ray Kurzweil) 现在是谷歌的工程总监，他于 1962 年在康奈尔大学的实验室遇到了罗森布拉特。库兹韦尔告诉我，他带了一些样品来试用感知器，只要字符以正确的字体清晰地打印出来，机器就能完美运行。罗森布拉特告诉当时即将进入麻省理工学院的年轻库兹韦尔，他相信如果感知器级联成多个级别，一个级别的输出作为下一个级别的输入，可以获得更好的结果。7^然而，罗森布拉特于 1971 年在一次划船事故中丧生，从未构建过多层实现。

The perceptron was capable of rudimentary pattern recognition tasks like recognizing printed characters via a camera that was attached to the device. The inventor and author Ray Kurzweil, who is now an engineering director at Google, met Rosenblatt in his lab at Cornell in 1962. Kurzweil told me that he brought samples to try out on the perceptron and that the machine worked perfectly as long as the characters were printed clearly in the proper font. Rosenblatt told the young Kurzweil, who was then about to matriculate at MIT, that he was confident that much better results could be obtained if perceptrons were cascaded into multiple levels, with the output of one level feeding into the inputs of the next.⁷ Rosenblatt, however, died in a boating accident in 1971, having never built a multilayer implementation.

到 20 世纪 60 年代末，人们对人工神经网络的热情开始消退。这种衰退最重要的驱动因素之一是 1969 年出版的《感知器》一书，由 Marvin Minsky 合著。尽管 Minsky 对整个人工智能的前景非常有信心，但具有讽刺意味的是，他对有朝一日会带来空前进步的具体方法却非常悲观。在书中，Minsky 和合著者 Seymour Papert 提出了正式的数学证明，强调了神经网络的局限性，并指出该技术将无法解决复杂的实际问题^。8

The initial enthusiasm for artificial neural networks began to drain away by the late 1960s. One of the most important drivers of this decline was the publication of the 1969 book Perceptrons, co-authored by Marvin Minsky. While Minsky was extraordinarily confident about the prospects for artificial intelligence as a whole, he was ironically very pessimistic about the specific approach that would one day lead to unprecedented progress. In the book, Minsky and co-author Seymour Papert presented formal mathematical proofs that highlighted the limitations of neural networks and suggested that the technology would prove unable to solve sophisticated practical problems.⁸

随着计算机科学家和研究生开始回避研究神经网络，符号人工智能方法（现在通常被称为“经典人工智能”）占据了主导地位。神经网络在 20 世纪 80 年代和 90 年代曾短暂复苏，但几十年来，符号学派一直占据主导地位，尽管整个人工智能领域的热情在两个极端之间循环。对于联结主义者来说，人工智能的寒冬来得可怕，持续时间也长得惊人，即使符号人工智能的实践者享受着温暖的春天，寒冬也常常持续。

As computer scientists and graduate students began to shy away from working on neural networks, the symbolic AI approach—now often referred to as “classical AI”—became dominant. Neural networks would see brief resurgences in the 1980s and again in the 1990s, but for decades the symbolic school would reign, even as enthusiasm for the field of artificial intelligence as a whole cycled between extremes. For the connectionists, AI winters unfolded with frightening severity and duration, often persisting even while the practitioners of symbolic AI enjoyed a balmy spring.

整个 20 世纪 70 年代和 80 年代初，这种冻结尤其严重。如今被认为是深度学习主要奠基人之一的 Yann LeCun 告诉我，在那段时期，对神经网络的研究“比边缘化更糟糕”，并且“你不能发表一篇提到‘神经网络’一词的论文，因为它会立即被拒绝。” ⁹尽管如此，一小部分研究人员仍然对联结主义的愿景充满信心。他们中的许多人不是计算机科学背景，而是心理学或人类认知背景，他们的兴趣是由创建大脑功能数学模型的愿望驱动的。20 世纪 80 年代初，加州大学圣地亚哥分校心理学教授 David Rumelhart 构思了一种称为“反向传播”的技术，它至今仍是多层神经网络中使用的主要学习算法。鲁梅尔哈特与东北大学计算机科学家罗纳德·威廉姆斯和当时在卡内基梅隆大学的杰弗里·辛顿一起，在 1986 年发表于《自然》杂志的一篇论文中描述了该算法的使用方法，该论文如今被认为是人工智能领域最重要的科学论文之一。10反向^传播代表了根本的概念突破，有朝一日深度学习将主导人工智能领域，但计算机的速度要快到足以真正利用这种方法还需要几十年的时间。1981 年，杰弗里·辛顿还是加州大学圣地亚哥分校的一名年轻博士后研究员，与鲁梅尔哈特共事，11^后来他或许成为深度学习革命中最杰出的人物。

Throughout the 1970s and early 1980s, the freeze was especially hard. Yann LeCun, who is now considered to be one of the primary architects of deep learning, told me that during this period, research into neural networks was “worse than marginalized” and that “you couldn’t publish a paper that even mentioned the phrase ‘neural networks’ because it would immediately be rejected.”⁹ Still, a small number of researchers maintained faith in the connectionist vision. Many of these individuals had backgrounds not in computer science but in psychology or human cognition, and their interest was driven by a desire to create a mathematical model for the function of the brain. In the early 1980s, David Rumelhart, a psychology professor at the University of California, San Diego, conceived the technique known as “backpropagation,” which is still the primary learning algorithm used in multilayered neural networks today. Rumelhart, along with Ronald Williams, a computer scientist at Northeastern University, and Geoffrey Hinton, then at Carnegie Mellon, described how the algorithm could be used in what is now considered to be one of the most important scientific papers in artificial intelligence, published in the journal Nature in 1986.¹⁰ Backpropagation represented the fundamental conceptual breakthrough that would someday lead deep learning to dominate the field of AI, but it would be decades before computers would become fast enough to truly leverage the approach. Geoffrey Hinton, who had been a young postdoctoral researcher working with Rumelhart at UC San Diego in 1981,¹¹ would go on to become perhaps the most prominent figure in the deep learning revolution.

到 20 世纪 80 年代末，神经网络的实际应用开始出现。当时 AT&T 贝尔实验室研究员 Yann LeCun 在一种名为“卷积神经网络”的新架构中使用了反向传播算法。在卷积网络中，人工神经元的连接方式受到哺乳动物大脑视觉皮层的启发，这些网络的设计尤其适用于图像识别。LeCun 的系统可以识别手写数字，到 20 世纪 90 年代末，卷积神经网络让 ATM 机能够识别银行支票上的数字。

By the end of the 1980s, practical applications for neural networks began to emerge. Yann LeCun, then a researcher at AT&T’s Bell Labs, used the backpropagation algorithm in a new architecture called a “convolutional neural network.” In convolutional networks, the artificial neurons are connected in a way that is inspired by the visual cortex in the brains of mammals, and these networks were designed to be especially effective at image recognition. LeCun’s system could recognize handwritten digits, and by the late 1990s convolutional neural networks were allowing ATM machines to understand the numbers written on bank checks.

21 世纪迎来了“大数据”的崛起。各组织和政府正在收集并试图分析信息，而就在不久前，这种规模还是难以想象的。显而易见，全球产生的数据总量将继续呈指数级增长。这种数据井喷很快将与最新的机器学习算法相交织，引发人工智能革命。

The 2000s saw the rise of “big data.” Organizations and governments were now collecting and attempting to analyze information at what would, just a short time ago, have been an unimaginable scale, and it had become evident that the total volume of data generated globally would continue to grow at an exponential pace. This data gusher would soon intersect with the latest machine learning algorithms to enable a revolution in artificial intelligence.

其中一个最重要的新数据库来自普林斯顿大学一位年轻计算机科学教授的努力。李飞飞的工作重点是计算机视觉，她意识到，要教机器从视觉上理解现实世界，需要一个全面的教学资源，其中包含正确标记的示例，展示人物、动物、建筑物、车辆、物体以及人们可能遇到的任何其他事物的多种变化。在两年半的时间里，她着手为 5,000 多个类别的 300 多万张图片命名。这项工作必须手动完成；只有人类才能将照片与描述性标签正确关联起来。由于雇用本科生来承担如此艰巨的任务的成本高得令人望而却步，李的团队转向了亚马逊的 Mechanical Turk，这是一个新开发的平台，将信息导向的任务众包给远程工作者，通常是在低工资国家^。12

One of the most consequential new data troves resulted from the efforts of a young computer science professor at Princeton University. Fei-Fei Li, whose work was focused on computer vision, realized that teaching machines to make visual sense of the real world would require a comprehensive teaching resource with properly labeled examples showing many variations of people, animals, buildings, vehicles, objects—and just about anything else one might encounter. Over a two-and-a-half-year period, she set out to give titles to more than three million images across over 5,000 categories. This work had to be done manually; only a human being could make the proper association between a photograph and a descriptive label. Because the cost of hiring even undergraduates to take on such a massive task would have been prohibitive, Li’s team turned to Amazon’s Mechanical Turk, a newly developed platform that crowdsourced information-oriented tasks to remote workers, often in low-wage countries.¹²

李飞飞的项目 ImageNet 于 2009 年发布，很快成为机器视觉研究不可或缺的资源。从 2010 年开始，李飞飞组织了一场年度竞赛，来自大学和企业研究实验室的团队在竞赛中自由发挥他们的算法，试图标记从海量数据集中提取的图像。两年后，即 2012 年 9 月举行的 ImageNet 大规模视觉识别竞赛可以说是深度学习技术的转折点。13^杰夫·辛顿 (Geoff Hinton) 与他在多伦多大学研究实验室的伊利亚·苏茨克弗 (Ilya Sutskever) 和亚历克斯·克里热夫斯基 (Alex Krizhevsky) 一起，提出了一个多层卷积神经网络，该网络的速度大大超越了竞争对手的算法，并提供了明确的证据，表明深度神经网络终于发展成为一种真正实用的技术。辛顿团队的胜利在人工智能研究界引起了广泛共鸣，并让人们关注到海量数据集与强大的神经算法的有效结合——这种共生关系很快就会产生几年前还只存在于科幻小说领域的进步。

Li’s project, known as ImageNet, was published in 2009 and soon became an indispensable resource for research into machine vision. Beginning in 2010, Li organized an annual competition in which teams from universities and corporate research labs turned their algorithms loose to try to label images drawn from the massive dataset. The ImageNet Large Scale Visual Recognition Competition held two years later, in September 2012, arguably represents the inflection point for the technology of deep learning.¹³ Geoff Hinton, along with Ilya Sutskever and Alex Krizhevsky from his research lab at the University of Toronto, entered a many-layered convolutional neural network that dramatically outpaced competing algorithms and offered unambiguous evidence that deep neural networks had finally evolved into a genuinely practical technology. The triumph of Hinton’s team resonated widely within the AI research community and put a spotlight on the productive coupling of massive datasets with powerful neural algorithms—a symbiosis that would soon produce advances that just a few years ago had been anchored firmly in the realm of science fiction.

我在这里勾勒出的故事大致代表了所谓的深度学习的“标准历史”。在这个故事中，2018 年图灵奖获得者 Geoff Hinton、Yann LeCun 和蒙特利尔大学教授 Yoshua Bengio 尤为重要——以至于他们经常被称为“深度学习教父”。（他们有时也被称为“人工智能教父”，生动地展示了深度学习已经完全主导了该领域的程度，将早期对符号方法的关注推到了一边。）然而，这段历史还有其他版本。与大多数科学领域一样，对认可的竞争非常激烈——也许因为人们越来越意识到人工智能的进步已经跨越了门槛，这将不可避免地导致社会和经济发生真正的历史性转变，这种竞争已经走向了极端。

The story I’ve sketched out here represents roughly what you might call the “standard history” of deep learning. In this telling, the 2018 Turing Award recipients Geoff Hinton, Yann LeCun and Yoshua Bengio, a professor at the University of Montreal, loom especially large—so much so that they are often referred to as the “godfathers of deep learning.” (They are also sometimes called “the godfathers of AI,” a vivid demonstration of the extent to which deep learning has come to completely dominate the field, pushing aside the earlier focus on symbolic approaches.) There are other versions of this history, however. As with most scientific fields, the competition for recognition is keen—and has perhaps been driven to extremes by the growing sense that progress in AI has crossed thresholds that will inevitably lead to genuinely historic transformations of both society and the economy.

最支持替代历史的人是瑞士卢加诺 Dalle Molle 人工智能研究所的共同主任 Jürgen Schmidhuber。20 世纪 90 年代，Schmidhuber 和他的学生开发了一种特殊的神经网络，实现了“长短期记忆”，即 LSTM。借助 LSTM，网络能够“记住”过去的数据并将其纳入当前分析。事实证明，这在语音识别和语言翻译等领域至关重要，因为先前单词所形成的上下文对准确性有很大影响。谷歌、亚马逊和 Facebook 等公司都严重依赖 LSTM，Schmidhuber 认为，人工智能近期取得的大部分进展是他的团队的工作成果，而不是更著名的北美研究人员的工作成果。

The most vocal proponent of an alternate history is Jürgen Schmidhuber, who co-directs the Dalle Molle Institute for Artificial Intelligence Research in Lugano, Switzerland. In the 1990s, Schmidhuber and his students developed a special type of neural network that implemented “long short-term memory,” or LSTM. With LSTM, networks are able to “remember” data from the past and incorporate it into the current analysis. This has proven to be of critical importance in areas like speech recognition and language translation, where the context created by words that came previously has a huge impact on accuracy. Companies like Google, Amazon and Facebook all rely heavily on LSTM, and Schmidhuber feels that it is the work of his team, rather than that of the more celebrated North American researchers, that underlies much of AI’s recent progress.

在我出版《智能建筑师》一书后不久，我给施密德胡贝发了一封电子邮件，其中我简要介绍了深度学习的标准历史。施密德胡贝告诉我：“你写的很多东西都具有误导性，因此非常令人沮丧！” ¹⁴施密德胡贝认为，深度学习的根源不是美国或加拿大，而是欧洲。他说，第一个多层神经网络的学习算法是由乌克兰研究员阿列克谢·格里戈列维奇·伊瓦赫年科 (Alexey Grigorevich Ivakhnenko) 于 1965 年描述的，而反向传播算法则由芬兰学生塞波·林奈因马 (Seppo Linnainmaa) 于 1970 年发表，比鲁梅尔哈特的著名论文早了十五年。施米德胡贝显然对自己的研究成果得不到认可感到沮丧，他以粗鲁地打断人工智能会议上的演讲和指责“阴谋”改写深度学习历史而闻名，尤其是辛顿、勒昆和本吉奥。15^而这些知名研究人员则积极反击。勒昆告诉《纽约时报》记者：“尤尔根疯狂地痴迷于认可，不断索要他不应得的荣誉。” ¹⁶

In an email written to me shortly after the publication of my book Architects of Intelligence—in which I included a short summary of deep learning’s standard history—Schmidhuber told me that “many of the things you wrote are quite misleading and, as a consequence, quite frustrating!”¹⁴ Deep learning, according to Schmidhuber, has its roots not in the United States or Canada but rather in Europe. The first learning algorithms for multilayered neural networks, he says, were described by the Ukrainian researcher Alexey Grigorevich Ivakhnenko in 1965, while the backpropagation algorithm was published in 1970—a decade and a half before the famous paper by Rumelhart—by Finnish student Seppo Linnainmaa. Schmidhuber is clearly frustrated over the lack of recognition given to his own research, and is known for abrasively interrupting presentations at AI conferences and leveling accusations of a “conspiracy” to rewrite deep learning’s history, especially on the part of Hinton, LeCun and Bengio.¹⁵ For their part, these better-known researchers push back aggressively. LeCun told a New York Times reporter that “Jürgen is manically obsessed with recognition and keeps claiming credit he doesn’t deserve.”¹⁶

尽管关于深度学习真正起源的争议可能仍会持续，但毫无疑问的是，在 2012 年 ImageNet 竞赛之后，这项技术迅速席卷了人工智能领域，并席卷了大多数科技行业的大公司。谷歌、亚马逊、Facebook 和苹果等美国科技巨头，以及百度、腾讯和阿里巴巴等中国公司，立即意识到深度神经网络的颠覆性潜力，并开始组建研究团队，将这项技术融入到他们的产品和运营中。谷歌聘请了杰夫·辛顿 (Geoff Hinton)，扬·勒昆 (Yann LeCun) 成为 Facebook 新人工智能研究实验室的主任，整个行业开始发动一场全面的人才争夺战，即使是刚毕业的、拥有深度学习专业知识的学生，他们的薪水和股票期权也飙升到天价。2017 年，首席执行官桑达尔·皮查伊 (Sundar Pichai) 宣布谷歌现在是一家“人工智能优先的公司”，并表示人工智能将成为该公司与其他科技巨头竞争的最重要方面之一。¹⁷在谷歌和 Facebook，这项技术被认为非常重要，以至于深度学习研究人员的办公室被安排在离首席执行官很近的地方，¹⁸到了 20 世纪 90 年代末，神经网络已经完全主宰了该领域，媒体经常将“深度学习”和“人工智能”视为同义词。

Though disagreements about the true origins of deep learning are likely to persist, there is no doubt that in the wake of the 2012 ImageNet competition, the technique rapidly took the field of artificial intelligence—and most of the technology industry’s largest companies—by storm. American tech behemoths like Google, Amazon, Facebook and Apple, as well as the Chinese companies Baidu, Tencent and Alibaba, immediately recognized the disruptive potential of deep neural networks and began to build research teams and incorporate the technology into their products and operations. Google hired Geoff Hinton, Yann LeCun became the director of Facebook’s new AI research lab, and the entire industry began waging a full-on talent war that pushed salaries and stock options for even newly minted graduate students with expertise in deep learning into the stratosphere. In 2017, CEO Sundar Pichai declared that Google was now an “AI-first company” and said that artificial intelligence would be one of the most important dimensions along which the company would compete with the other tech giants.¹⁷ At Google and Facebook, the technology was deemed so important that deep learning researchers were assigned offices in direct proximity to the CEO,¹⁸ and by the end of the decade, neural networks had so completely dominated the field that the media would often treat the terms “deep learning” and “artificial intelligence” as synonymous.

¹⁸

第五章

CHAPTER 5 深度学习与人工智能的未来

DEEP LEARNING AND THE FUTURE OF ARTIFICIAL INTELLIGENCE

全球最大的科技公司纷纷拥抱深度学习，再加上越来越多利用神经网络力量的引人注目的消费者和商业应用程序的出现，毫无疑问，这项技术将会继续存在。然而，越来越多的人认为，这种发展速度是不可持续的，未来的进步将需要重大的创新。正如我们将看到的，未来最重要的问题之一将是人工智能钟摆是否会再次摆回强调符号人工智能的方法，如果会，这些想法如何成功地与神经网络相结合。在深入探讨人工智能的未来之前，让我们先简要详细地了解一下深度学习系统的实际工作原理，以及如何训练这些网络来执行有用的任务。

THE EMBRACE OF DEEP LEARNING BY THE WORLD’S LARGEST technology companies, together with the arrival of ever more compelling consumer and business applications that leverage the power of neural networks, leaves little doubt that the technology is here to stay. However, there is a growing sense that the rate of progress is not sustainable, and that future advances will require significant new innovations. As we’ll see, one of the most important questions going forward will be whether or not the AI pendulum will once again swing back toward approaches that emphasize symbolic AI and, if so, how those ideas can be successfully integrated with neural networks. Before delving into the future of artificial intelligence, let’s briefly look in more detail at just how a deep learning system actually works and how these networks are trained to perform useful tasks.

深度神经网络的工作原理

HOW DEEP NEURAL NETWORKS WORK

媒体经常将深度学习系统描述为“类似大脑”，这很容易导致人们对人工智能中使用的神经网络与生物神经网络的相似程度产生误解。人类大脑可以说是已知宇宙中最复杂的系统，拥有大约一千亿个神经元和数百万亿个互连。但这种惊人的复杂性并非仅仅源于大规模的连接；相反，它延伸到神经元本身的运作以及它们传输信号和随时间适应新信息的方式。

The media often describes deep learning systems as “brain-like,” and this can easily lead to misconceptions about just how closely the neural networks used in artificial intelligence are modeled on their biological counterparts. The human brain is arguably the most complex system in the known universe, with roughly one hundred billion neurons and hundreds of trillions of interconnections. But this staggering level of complexity does not arise simply from connectivity at a massive scale; rather, it extends to the operation of the neurons themselves and the way they transmit signals and adapt to new information over time.

生物神经元由三个主要部分组成：细胞体，细胞核位于其中；无数的细丝，称为“树突”，用于传输传入的电信号；以及一条更长更细的细丝，称为“轴突”，神经元通过它向其他神经元传输传出信号。树突和轴突通常都有广泛的分支，树突有时会从成千上万个其他神经元接收电刺激。当通过树突到达的集体信号刺激神经元时，它反过来通过轴突传递一个传出的电荷，称为动作电位。然而，大脑的连接并不是电硬连线的。相反，一个神经元的轴突通过一个称为“突触”的连接点将化学信号传递给另一个神经元的树突。这些电化学功能对大脑的运作及其学习和适应能力至关重要，但在很多情况下，人们对其了解甚少。例如，与愉悦感或奖励有关的化学物质多巴胺是一种在突触间隙内起作用的神经递质。

A biological neuron consists of three main parts: the cell body, where the nucleus resides; numerous filaments known as “dendrites” that carry incoming electrical signals; and a single, much longer and finer filament called an “axon,” along which the neuron transmits an outgoing signal to other neurons. Both the dendrites and the axon typically branch extensively, with the dendrites sometimes receiving electrical stimulation from tens of thousands of other neurons. When the collective signals arriving through the dendrites excite the neuron, it in turn delivers an outgoing electrical charge, known as an action potential, through the axon. However, the brain’s connections are not electrically hardwired. Instead, the axon from one neuron transmits a chemical signal to a dendrite of another across a junction known as a “synapse.” These electrochemical functions are critical to the brain’s operation and to its ability to learn and adapt, but are, in many cases, not well understood. The chemical dopamine, for example, which is associated with pleasure or reward, is a neurotransmitter that operates within the synaptic gap.

人工神经网络几乎忽略了所有这些细节，试图对神经元的运作和连接方式进行粗略的数学描述。如果说大脑是蒙娜丽莎，那么深度学习系统中使用的结构充其量也只是类似于《花生漫画》中的露西。人工神经元的基本方案是在 20 世纪 40 年代构思出来的，在此后的几十年里，这些系统的研究基本上与脑科学脱钩；深度学习系统所依赖的算法都是独立开发的，通常是通过实验开发的，并没有特别尝试模拟人类大脑中可能发生的事情。

An artificial neural network brushes aside nearly all these details and attempts to create a rough mathematical sketch of the way neurons operate and connect. If the brain is the Mona Lisa, then the structures used in deep learning systems might, at best, be something akin to Lucy from Peanuts. The basic plan for artificial neurons was conceived in the 1940s, and in the decades since, work on these systems has largely been decoupled from brain science; the algorithms that power deep learning systems have been developed independently, often through experimentation and without any specific attempt to simulate what might actually be occurring in the human brain.

为了形象化地展示人造神经元，想象一个容器，里面有三根或更多根进水管，每根管子都输送一股水流。这些管子大致相当于生物神经元中的树突。还有一个轴突管，用来输送一股流出的水流。如果进水管输送的水量达到某个阈值，神经元就会“激发”，然后通过轴突管输送一股流出的水流。

To visualize an artificial neuron, imagine a container with three or more incoming pipes, each of which delivers a stream of water. These pipes correspond roughly to dendrites in the biological neuron. There is also an axon pipe to carry an outgoing stream of water. If the level of water delivered by the incoming pipes reaches a certain threshold, the neuron “fires” then sends an outgoing stream of water through its axon pipe.

将这个简单的装置变成有用的计算设备的关键特征是安装在每个进水管上的阀门，这样就可以控制通过管道的水流。通过调节这些阀门，可以直接调节其他连接的神经元对这个特定神经元的影响。训练神经网络执行一些有用任务的过程本质上就是调节这些阀门（称为“权重”），直到网络能够正确识别模式。

The key feature that turns this simple contraption into a useful computational device is a valve that is fitted to each of the incoming pipes, so that the flow of water through the pipe can be controlled. By adjusting these valves, it’s possible to directly regulate the influence that other connected neurons have on this particular neuron. The process of training a neural network to perform some useful task is essentially a matter of adjusting these valves—called “weights”—until the network can correctly identify patterns.

在深度神经网络中，人工神经元的软件模拟工作方式与这些容器大致相同，它们会被排列成一系列层，这样一层神经元的输出就会连接到下一层神经元的输入。相邻层神经元之间的连接通常只是随机设置的；或者，在特定的神经结构中，例如用于识别图像的卷积网络，神经元可能会根据更精心设计的计划进行连接。复杂的神经网络可以包含一百多层和数百万个单独的人工神经元。

In a deep neural network, a software simulation of artificial neurons that work more or less like these containers would be arranged in a series of layers, so that the output from one layer of neurons is connected to the inputs of neurons in the next layer. The connections between neurons in adjacent layers are often simply set randomly; alternatively, in a specific neural architecture, such as a convolutional network designed to recognize images, the neurons may be connected according to a more deliberate plan. Sophisticated neural networks can contain more than a hundred layers and millions of individual artificial neurons.

一旦配置好这样的网络，就可以训练它执行特定任务，例如图像识别或语言翻译。例如，为了训练神经网络识别手写数字，手写数字照片中的各个像素将成为第一层神经元的输入。答案，或者换句话说，与手写数字相对应的数字，将到达最后一层人工神经元的输出。训练网络生成正确答案是一个输入训练示例然后调整网络中的所有权重的过程，以便它逐渐收敛到正确答案。一旦以这种方式优化了权重，网络就可以部署在训练集中未包含的新示例上。

Once such a network has been configured, it can then be trained to perform specific tasks, such as image recognition or language translation. For example, in order to train a neural network to recognize handwritten digits, the individual pixels from a photograph of a written numeral would become the inputs for the first layer of neurons. The answer, or in other words the number corresponding to the written digit, would arrive at the outputs from the last layer of artificial neurons. Training the network to generate the correct answer is a process of inputting training examples and then adjusting all the weights in the network so that it gradually converges on the right answer. Once the weights have been optimized in this way, the network can then be deployed on new examples that are not included in the training set.

调整权重，使网络几乎每次都能成功收敛到正确答案，这就是著名的反向传播算法的作用所在。一个复杂的深度学习系统可能有十亿个或更多的神经元连接，每个神经元都有一个需要优化的权重。反向传播本质上允许集体调整网络中的所有权重，而不是一次调整一个，从而大大提高计算效率。1^在训练过程中，网络的输出与正确答案进行比较，允许相应调整每个权重的信息会通过神经元层反向传播。没有反向传播，深度学习革命就不可能实现。

Tuning the weights so the network eventually succeeds in converging on the right answer nearly every time is where the famous backpropagation algorithm comes in. A complex deep learning system might have a billion or more connections between neurons, each of which has a weight that needs to be optimized. Backpropagation essentially allows all the weights in the network to be adjusted collectively, rather than one at a time, delivering a massive boost to computational efficiency.¹ During the training process, the output from the network is compared to the correct answer, and information that allows each weight to be adjusted accordingly propagates back through the layers of neurons. Without backpropagation, the deep learning revolution would not have been possible.

虽然所有这些都勾勒出了配置和训练神经网络以产生有用结果的基本机制，但它仍然没有回答一个基本问题：当这些系统处理数据并以超人的准确度提供答案时，其中究竟发生了什么？

While all this sketches out the basic mechanics of configuring and training a neural network so that it will produce useful results, it still leaves unanswered the fundamental question: what exactly is actually happening within one of these systems as it churns through data and delivers answers with an accuracy that is often superhuman?

简而言之，在神经网络中，知识的表示正在被创建，并且这种知识的抽象级别在网络的后续层中不断增加。对于配置为识别视觉图像的网络来说，这是最容易理解的。网络对图像的理解始于像素级别。在后续的神经层中，可以感知边缘、曲线和纹理等视觉特征。在系统更深处，会出现更复杂的表示。最终，系统的理解是如此明确，以至于它能够以一种允许网络识别它的方式捕捉图像的全部本质——即使面对大量替代方案。

The short explanation is that, within the neural network, a representation of knowledge is being created, and the level of abstraction for this knowledge increases in subsequent layers of the network. This is easiest to understand for a network configured to recognize visual images. The network’s comprehension of the image begins at the level of pixels. In subsequent neural layers, visual features such as edges, curves and textures are perceived. Deeper within the system, still more complex representations emerge. Eventually, the system’s understanding is so definitive that it captures the full essence of the image in a way that allows the network to identify it—even when confronted with a huge number of alternatives.

然而，对这个问题更完整的回答是承认我们并不真正知道到底发生了什么——或者至少我们不能轻易地描述它。没有程序员会着手定义各种抽象层次或知识在网络中的表示方式。所有这些都是有机地出现的，并且表示分布在整个系统中数百万个相互连接的人工神经元中。我们知道网络在某种意义上理解图像，但要准确描述其神经元内究竟在凝聚什么却非常困难甚至不可能——而且，当我们深入网络层时，或者当我们检查操作不那么容易可视化的数据类型的系统时，这种情况会变得越来越严重。这种相对不透明性——人们担心深度神经网络实际上是“黑匣子”——是最重要的问题之一，我们将在第 8 章中回顾这一点。

A more complete answer to the question, however, would admit we don’t really know exactly what’s happening—or at least we can’t readily describe it. No programmer sets out to define the various levels of abstraction or the way in which knowledge is represented within the network. All of this emerges organically, and the representation is distributed across the millions of interconnected artificial neurons firing throughout the system. We know that the network in some sense comprehends the image, and yet it is very difficult, or even impossible, to accurately describe just what is coalescing within its neurons—and this becomes ever more the case as we progress deeper into the layers of the network, or if we examine systems that operate on types of data that are not so easily visualized. This relative opacity—the concern that deep neural networks in effect are “black boxes”—is one of the most important concerns that we will return to in Chapter 8.

绝大多数深度学习系统都是通过向网络提供经过仔细标记或分类的海量数据集来训练以执行有用的任务的。例如，通过提供数千张甚至数百万张图像（每张图像都正确标记了所描绘动物的名称），可以训练深度神经网络正确识别照片中的动物。这种训练方案称为“监督学习”，即使使用非常高性能的硬件，也可能需要花费数小时。

The overwhelming majority of deep learning systems are trained to do useful tasks by presenting the network with a massive dataset that has been carefully labeled or categorized. For example, a deep neural network might be trained to correctly identify animals in photographs by being provided thousands, or perhaps even millions, of images, each correctly labeled with the name of the animal depicted. This training regimen, known as “supervised learning,” can take many hours even when very high performance hardware is used.

监督学习是大约 95% 的实际机器学习应用中使用的训练方法。该技术为人工智能放射系统（使用大量标记为“癌症”或“无癌症”的医学图像进行训练）、语言翻译（使用数百万份预先翻译成不同语言的文档进行训练）以及几乎无限数量的其他应用程序提供支持，这些应用程序本质上涉及比较和分类不同形式的信息。监督学习通常需要大量标记数据，但结果可能非常令人印象深刻——通常会产生具有超人识别模式能力的系统。2012 年 ImageNet 竞赛标志着深度学习爆炸式增长的开始，五年后，图像识别算法变得如此熟练，以至于年度竞赛重新定位为涉及识别现实世界三维物体的新挑战^。2

Supervised learning is the training method used in perhaps ninety-five percent of practical machine learning applications. The technique powers AI radiology systems (trained with a huge number of medical images labeled either “Cancer” or “No Cancer”), language translation (trained with millions of documents pre-translated into different languages) and a nearly limitless number of other applications that essentially involve comparing and classifying different forms of information. Supervised learning typically requires vast amounts of labeled data, but the results can be very impressive—routinely resulting in systems with a superhuman ability to recognize patterns. Five years after the 2012 ImageNet competition that marked the onset of the deep learning explosion, the image recognition algorithms had become so proficient that the annual competition was reoriented toward a new challenge involving the recognition of real-world three-dimensional objects.²

如果标记所有这些数据需要只有人类才能提供的解释，例如在照片上附加描述性注释，那么这个过程既昂贵又繁琐。一种常见的解决方案是复制李飞飞为 ImageNet 数据集采用的方法，并转向众包。Mechanical Turk 等平台可以以极低的价格向分布式团队支付这项工作。简化这一过程的机会催生了许多初创公司，它们专注于寻找有效的数据注释方法，为监督学习做准备。准确标记海量数据集至关重要，尤其是对于涉及理解视觉信息的应用程序，Scale AI 的迅速崛起尤其充分地证明了这一点，该公司由 19 岁的麻省理工学院辍学生 Alexandr Wang 于 2016 年创立。Scale AI 与 30,000 多名众包工人签订了合同，他们为 Uber、Lyft、Airbnb 和 Alphabet 的自动驾驶汽车部门 Waymo 等客户标记数据。该公司已获得超过 1 亿美元的风险投资，目前已跻身硅谷“独角兽”行列——估值超过 10 亿美元的初创公司^。3

In cases where labeling all this data requires the kind of interpretation that only a human can provide, as in attaching descriptive annotations to photographs, the process is expensive and cumbersome. One common solution is to copy the approach employed by Fei-Fei Li for the ImageNet dataset and turn to crowdsourcing. Platforms like Mechanical Turk make it possible to pay a distributed team of people pennies to do this work. The opportunity to streamline this process has given rise to a number of startup companies that are specifically focused on finding efficient ways to annotate data in preparation for supervised learning. The critical importance of accurately labeling massive datasets, especially for applications that involve understanding visual information, is especially well demonstrated by the meteoric ascent of Scale AI, which was founded by nineteen-year-old MIT dropout Alexandr Wang in 2016. Scale AI contracts with over 30,000 crowdsourced workers who label data for clients including Uber, Lyft, Airbnb and Alphabet’s self-driving car division, Waymo. The company has received more than $100 million in venture capital and now ranks as a Silicon Valley “unicorn”—a startup valued in excess of $1 billion.³

然而，在许多其他情况下，几乎难以想象的大量标记精美的数据似乎是自动生成的——对于拥有这些数据的公司来说，这几乎是免费的。Facebook、谷歌或 Twitter 等平台生成的大量数据之所以有价值，很大程度上是因为这些数据都经过了平台用户的仔细注释。每次你“点赞”或“转发”帖子、每次你查看或向下滚动网页、每次你观看视频（以及你观看视频所花费的时间）以及每次你进行无数其他在线操作时，你实际上都会为某个特定的数据项贴上标签。你——以及使用其中一个主要平台的数百万其他人——实际上是在接替 Scale AI 等公司部署的所有众包工作者。当然，最重要的人工智能研究计划往往与大型互联网公司有关，这并非巧合。人们经常谈论人工智能与海量数据所有权之间的协同作用，但这种共生关系背后的一个关键因素是拥有一台巨大的机器，可以以极低的成本或免费注释所有数据，因此它可以成为部署在强大神经网络上的监督学习方案的素材。

In many other cases, however, nearly incomprehensible quantities of beautifully labeled data are generated seemingly automatically—and for the companies that possess it, virtually free of charge. The massive torrent of data generated by platforms like Facebook, Google or Twitter is valuable in large measure because it is carefully annotated by the people using the platforms. Every time you “like” or “retweet” a post, every time you view or scroll down a webpage, every video you watch (and the amount of time you spend watching it) and every time you undertake a myriad of other online actions, you in effect attach a label to some particular item of data. You—along with the millions of other people using one of the major platforms—are essentially stepping into the shoes of all those crowdsourced workers deployed by companies like Scale AI. It is, of course, not coincidental that most important AI research initiatives tend to be associated with large internet companies. The synergy between artificial intelligence and ownership of huge troves of data is often remarked on, but a critical factor underlying this symbiosis is the possession of a massive machine for annotating all that data at little or no cost, so it can then become fodder for a supervised learning regimen deployed on a powerful neural network.

虽然监督学习占主导地位，但另一项重要技术“强化学习”也用于某些应用。强化学习通过反复练习或反复试验来建立能力。当算法最终成功实现指定目标时，它会获得数字奖励。这基本上就是训练狗的方式。动物的行为一开始可能是随机的，但当它设法响应正确的命令坐下时，它会得到奖励。重复这个过程足够多次，狗就会学会可靠地坐下。

While supervised learning dominates, another important technique—“reinforcement learning”—is used in certain applications. Reinforcement learning builds competence through repeated practice or trial and error. When an algorithm ultimately succeeds at a specified objective, it receives a digital reward. This is essentially the way a dog is trained. The animal’s behavior may be random at first, but when it manages to sit in response to the proper command, it gets a treat. Repeat this process enough times and the dog will learn to reliably sit.

强化学习领域的领导者是总部位于伦敦的 DeepMind 公司，该公司目前归谷歌母公司 Alphabet 所有。DeepMind 大规模投资于基于该技术的研究，将其与强大的卷积神经网络相结合，开发出公司所称的“深度强化学习”。2010 年成立后不久，DeepMind 就开始致力于应用强化学习来构建可以玩视频游戏的人工智能系统。2013 年 1 月，该公司宣布已创建一个名为 DQN 的系统，该系统能够玩经典的 Atari 游戏，包括《太空侵略者》、《Pong》和《Breakout》。DeepMind 的系统能够仅使用原始像素和游戏分数作为学习输入，来自学玩游戏。在数千场模拟游戏中磨练其技术后，DQN 在六场游戏中创下了计算机有史以来的最高分，并在三场游戏中击败了最优秀的人类玩家。⁴到 2015 年，该系统已征服了 49 款 Atari 游戏，DeepMind 宣称它已开发出第一个弥合“高维感官输入和动作之间的鸿沟”的人工智能系统，并且 DQN“能够学习并擅长处理各种具有挑战性的任务。” ⁵这些成就引起了硅谷巨头的注意，尤其是谷歌创始人拉里·佩奇，2014 年，谷歌拒绝了 Facebook 的竞争报价，以 4 亿美元收购了 DeepMind。

The leader in reinforcement learning is the London-based company DeepMind, which is now owned by Google’s parent, Alphabet. DeepMind has made massive investments in research based on the technique, merging it with powerful convolutional neural networks to develop what the company calls “deep reinforcement learning.” DeepMind began working on applying reinforcement learning to build AI systems that could play video games shortly after its founding in 2010. In January 2013, the company announced that it had created a system called DQN that was capable of playing classic Atari games, including Space Invaders, Pong and Breakout. DeepMind’s system was able to teach itself to play the games by using only raw pixels and the game score as the learning inputs. After honing its technique on many thousands of simulated games, DQN turned in the best scores ever achieved by a computer for six of the games, and was able to defeat the best human players in three.⁴ By 2015, the system had conquered forty-nine Atari games, and DeepMind declared that it had developed the first AI system that bridged “the divide between high-dimensional sensory inputs and actions” and that the DQN was “capable of learning to excel at a diverse array of challenging tasks.”⁵ These achievements caught the attention of Silicon Valley’s titans, including especially Google founder Larry Page, and in 2014 Google brushed aside a competing offer from Facebook and acquired DeepMind for $400 million.

深度强化学习最引人注目的成就出现在 2016 年 3 月，当时 DeepMind 开发的用于玩古老围棋游戏的系统 AlphaGo 在韩国首尔的一场五局比赛中击败了当时世界顶级围棋选手李世石。围棋在亚洲享有极高的声誉，因为围棋在当地已有数千年的历史。孔子的著作中曾提到过围棋，而它的起源可能可以追溯到中华文明的黎明时期。一种理论认为，围棋是在尧统治时期发明的，大约在公元前 2000 年之前。6^下围棋的能力与书法、绘画和弹奏弦乐器的技能一起，被视为中国古代学术的四大主要艺术之一。

Deep reinforcement learning’s most notable achievement came in March 2016, when AlphaGo, a system developed by DeepMind to play the ancient game of Go, defeated Lee Sedol, then one of the world’s best players, in a five-game match in Seoul, South Korea. Proficiency at Go is held in extremely high regard in Asia, where the game has been played for millennia. Confucius’s writings reference the game, and its origins may stretch back nearly to the dawn of Chinese civilization. According to one theory, Go was invented during the reign of the emperor Yao, sometime prior to 2000 BC.⁶ The ability to play Go, along with expertise in calligraphy, painting and playing a stringed musical instrument, was viewed as one of the four primary arts that marked ancient Chinese scholarship.

与国际象棋不同，围棋游戏非常复杂，以至于无法受到强力算法的攻击。在游戏过程中，棋盘由一个 19×19 的网格组成，其中大部分都是黑白棋子，称为“棋子”。正如 DeepMind 首席执行官 Demis Hassabis 在讨论 AlphaGo 的成就时经常喜欢指出的那样，棋盘上棋子可能的排列方式数量超过了已知宇宙中原子的数量。在围棋游戏的数千年历史中，任何两场游戏以相同方式展开的可能性极小——实际上几乎为零。换句话说，任何试图展望未来并考虑所有可能的未来走法范围的尝试（就像在规则更严格的游戏中所做的那样）在计算上都是遥不可及的——即使是最强大的硬件也无法做到。

Unlike chess, the game of Go is so complex as to be immune to the onslaught of brute force algorithms. During the course of the game, the board, which consists of a nineteen-by-nineteen grid, is largely filled with black and white game pieces called “stones.” As DeepMind CEO Demis Hassabis often likes to point out when he discusses AlphaGo’s accomplishment, the number of possible arrangements of the stones on the board exceeds the estimated number of atoms in the known universe. Over the thousands of years that the game has been played, it is extraordinarily—indeed vanishingly—unlikely that any two games have unfolded in identical fashion. In other words, any attempt to look ahead and account for the entire range of possible future moves, as might be done for a game with more confining rules, is computationally out of reach—even for the most powerful hardware.

除了这种巨大的复杂性之外，下围棋显然在很大程度上依赖于所谓的人类直觉。当被要求解释他们选择特定策略的确切原因时，最优秀的棋手往往不知所措。相反，他们可能会描述一种“感觉”，这种感觉促使他们将棋子放在棋盘上的某个位置。这种任务似乎超出了计算机的能力范围——我们可以理所当然地认为，这项工作至少在可预见的未来不会受到自动化的威胁。尽管如此，围棋游戏至少在大多数计算机科学家相信这一壮举可能实现的十年前就落入了机器的手中。

Aside from this vast level of complexity, it seems evident that playing Go draws heavily on what might be called human intuition. The best players are often at a loss when they are asked to explain exactly why they chose a particular strategy. Instead, they might describe a “feeling” that led them to place a stone at a particular point on the board. This is precisely the type of undertaking that seems as though it ought to be beyond the capability of a computer—a job that we can rightfully expect to be safe from the threat of automation, at least for the foreseeable future. Nonetheless, the game of Go fell to the machines at least a decade before most computer scientists believed such a feat would be possible.

DeepMind 团队首先使用监督学习技术训练 AlphaGo 的神经网络，这些神经网络是从人类顶尖棋手的详细比赛记录中提取的三千万步棋。然后，他们转向强化学习，让系统自由地与自己对弈。在数千场模拟练习赛中，在基于奖励的不断压力下，AlphaGo 的深度神经网络逐渐发展到超越人类的水平。7 ^AlphaGo于 2016 年战胜李世石，一年后又战胜世界排名第一的棋手柯洁，再次震惊了人工智能研究界。这些成就可能也导致了风险投资家兼作家李开复所说的中国“斯普特尼克时刻”——在此之后，政府迅速采取行动，将中国定位为人工智能领域的领导者^。8

The DeepMind team began by using a supervised learning technique to train AlphaGo’s neural networks on thirty million moves extracted from detailed records of games played by the best human players. It then turned to reinforcement learning, essentially turning the system loose to play against itself. Over the course of thousands of simulated practice games, and under the relentless pressure of a reward-based drive to improve, AlphaGo’s deep neural networks gradually progressed toward superhuman proficiency.⁷ The triumph of AlphaGo over Lee Sedol in 2016, and then over the world’s top-ranked player, Ke Jie, a year later, once again sent shock waves through the AI research community. The achievements may also have led to what venture capitalist and author Kai-Fu Lee has called a “Sputnik moment” in China—in the wake of which the government quickly moved to position the country to become a leader in artificial intelligence.⁸

监督学习依赖于大量标记数据，而强化学习则需要大量的练习，而大多数练习都以失败告终。强化学习特别适合游戏——在游戏中，算法可以快速处理比人类一生中玩的比赛还要多的比赛。这种方法还可以应用于可以高速模拟的现实世界活动。目前，强化学习最重要的实际应用是训练自动驾驶汽车。在 Waymo 或特斯拉使用的自动驾驶系统出现在真正的汽车或真正的道路上之前，它们会在强大的计算机上进行高速训练，通过这些训练，模拟汽车在遭遇数千次灾难性的碰撞后逐渐学习。一旦算法训练完成，碰撞成为过去，软件就可以转移到现实世界的汽车上。虽然这种方法通常是有效的，但不言而喻的是，没有一个 16 岁考驾照的孩子需要经历一千次碰撞才能弄清楚如何驾驶。机器学习的方式与人类大脑中利用极少数据进行学习的方式形成了鲜明对比，这既凸显了当前人工智能系统的局限性，也凸显了该技术未来巨大的改进潜力。

While supervised learning depends on massive quantities of labeled data, reinforcement learning requires a huge number of practice runs, the majority of which end in spectacular failure. Reinforcement learning is especially well suited to games—in which an algorithm can rapidly churn though more matches than a human being could play in a lifetime. The approach can also be applied to real-world activities that can be simulated at high speed. The most important practical application of reinforcement learning is currently in training self-driving cars. Before the autonomous driving systems used by Waymo or Tesla ever see a real car or a real road, they are trained at high speed on powerful computers, through which the simulated cars gradually learn after suffering catastrophic crashes thousands of times. Once the algorithms are trained and the crashes are in the past, the software can then be transferred to real-world cars. Though this approach is generally effective, it goes without saying that no license-seeking sixteen year old needs to crash a thousand times before figuring out how to drive. This stark contrast between how learning works in machines and how it operates with vastly less data in the human brain highlights both the limitations of current AI systems and the enormous potential for the technology to improve going forward.

警告信号

WARNING SIGNS

2010 年代可以说是人工智能历史上最激动人心、影响最深远的十年。虽然人工智能算法在概念上确实有所改进，但所有这些进步的主要驱动力只是在速度越来越快的计算机硬件上部署更广泛的深度神经网络，从而可以吸收越来越多的训练数据。自 2012 年引发深度学习革命的 ImageNet 竞赛以来，这种“扩展”策略就一直很明确。同年 11 月，《纽约时报》头版文章在更广泛的公众领域引起了人们对深度学习技术的关注。这篇由记者 John Markoff 撰写的文章以 Geoff Hinton 的一句话结尾：“这种方法的优点在于它可以完美地扩展。基本上你只需要继续让它变得更大、更快，它就会变得更好。现在已经没有回头路了。” ⁹

The 2010s were arguably the most exciting and consequential decade in the history of artificial intelligence. Though there have certainly been conceptual improvements in the algorithms used in AI, the primary driver of all this progress has simply been deploying more expansive deep neural networks on ever faster computer hardware where they can hoover up greater and greater quantities of training data. This “scaling” strategy has been explicit since the 2012 ImageNet competition that set off the deep learning revolution. In November of that year, a front-page New York Times article was instrumental in bringing awareness of deep learning technology to the broader public sphere. The article, written by reporter John Markoff, ends with a quote from Geoff Hinton: “The point about this approach is that it scales beautifully. Basically you just need to keep making it bigger and faster, and it will get better. There’s no looking back now.”⁹

然而，越来越多的证据表明，这一主要进步引擎正开始失去动力。根据研究机构 OpenAI 的一项分析，尖端人工智能项目所需的计算资源正在“呈指数级增长”，大约每 3.4 个月翻一番。10在 2019 年 12 月的《连线》杂志采访中，Facebook 人工智能副总裁杰罗姆·佩森蒂 (Jerome Pesenti) 表示，即使对于像 Facebook 这样财力雄厚的公司来说，这种情况在财务上也是不可持续的^：

There is increasing evidence, however, that this primary engine of progress is beginning to sputter out. According to one analysis by the research organization OpenAI, the computational resources required for cutting-edge AI projects is “increasing exponentially” and doubling about every 3.4 months.¹⁰ In a December 2019 Wired magazine interview, Jerome Pesenti, Facebook’s Vice President of AI, suggested that even for a company with pockets as deep as Facebook’s, this would be financially unsustainable:

当你扩展深度学习时，它的表现会更好，能够以更好的方式解决更广泛的任务。因此，扩展是有优势的。但显然这种进步速度是不可持续的。如果你看看顶级实验，你会发现每年的成本都在上涨 10 倍。现在，一个实验的成本可能是七位数，但不会涨到九位数或十位数，这是不可能的，没有人能负担得起^。11

When you scale deep learning, it tends to behave better and to be able to solve a broader task in a better way. So, there’s an advantage to scaling. But clearly the rate of progress is not sustainable. If you look at top experiments, each year the cost [is] going up 10-fold. Right now, an experiment might be in seven figures, but it’s not going to go to nine or ten figures, it’s not possible, nobody can afford that.¹¹

Pesenti 还对扩展继续成为进步的主要驱动力的可能性发出了严厉警告：“在某个时候，我们会遇到瓶颈。在很多方面，我们已经遇到了瓶颈。”除了扩展到更大的神经网络的财务限制之外，还有重要的环境考虑因素。马萨诸塞大学阿默斯特分校的研究人员在 2019 年进行的一项分析发现，训练一个非常大的深度学习系统可能会排放相当于五辆汽车在其整个运行寿命内的二氧化碳^。12

Pesenti goes on to offer a stark warning about the potential for scaling to continue to be the primary driver of progress: “At some point we’re going to hit the wall. In many ways we already have.” Beyond the financial limits of scaling to ever larger neural networks, there are also important environmental considerations. A 2019 analysis by researchers at the University of Massachusetts, Amherst, found that training a very large deep learning system could potentially emit as much carbon dioxide as five cars over their full operational lifetimes.¹²

即使可以克服财务和环境影响挑战（或许通过开发效率极高的硬件或软件），但仅将扩展作为一种策略也可能不足以取得持续进展。不断增加的计算投资已经产生了在狭窄领域具有非凡能力的系统，但越来越明显的是，深度神经网络受到可靠性限制，除非取得重要的概念突破，否则这可能使该技术不适合许多关键任务应用。该技术弱点最显著的证明之一是，Vicarious 的一组研究人员（我们在第 3 章 ^中遇到的一家专注于制造灵巧机器人的小公司）对 DeepMind 的 DQN 中使用的神经网络进行了分析，该系统已经学会了在 Atari 视频游戏中占据主导地位。13其中一项测试是在Breakout上进行的，在 Breakout 游戏中，玩家必须操纵桨来拦截快速移动的球。当屏幕上的球拍仅仅向上移动几个像素时——人类玩家甚至可能注意不到这种变化——系统之前超人的表现就会立即一落千丈。DeepMind 的软件无法适应这种微小的改变。要恢复到顶级性能，唯一的办法就是从头开始，用基于新屏幕配置的数据对系统进行彻底的重新训练。

Even if the financial and environmental impact challenges can be overcome—perhaps through the development of vastly more efficient hardware or software—scaling as a strategy simply may not be sufficient to produce sustained progress. Ever-increasing investments in computation have produced systems with extraordinary proficiency in narrow domains, but it is becoming increasingly clear that deep neural networks are subject to reliability limitations that may make the technology unsuitable for many mission critical applications unless important conceptual breakthroughs are made. One of the most notable demonstrations of the technology’s weaknesses came when a group of researchers at Vicarious—the small company focused on building dexterous robots that we met in Chapter 3—performed an analysis of the neural network used in DeepMind’s DQN, the system that had learned to dominate Atari video games.¹³ One test was performed on Breakout, a game in which the player has to manipulate a paddle to intercept a fast-moving ball. When the paddle was shifted just a few pixels higher on the screen—a change that might not even be noticed by a human player—the system’s previously superhuman performance immediately took a nose dive. DeepMind’s software had no ability to adapt to even this small alteration. The only way to get back to top-level performance would have been to start from scratch and completely retrain the system with data based on the new screen configuration.

这告诉我们，尽管 DeepMind 强大的神经网络确实实例化了Breakout屏幕的表示，但即使在网络深处的更高抽象层，这种表示仍然牢牢地锚定在原始像素上。显然，人们并没有将桨理解为可以移动的实际物体。换句话说，对于屏幕上的像素所代表的物质对象或控制其运动的物理原理，人们的理解还远远不够。它只是像素而已。虽然一些人工智能研究人员可能仍然相信，只要有更多的人工神经元层、在更快的硬件上运行并消耗更多的数据，最终可能会出现更全面的理解，但我认为这不太可能。在我们开始看到具有更像人类的世界观的机器之前，还需要更多根本性的创新。

What this tells us is that while DeepMind’s powerful neural networks do instantiate a representation of the Breakout screen, this representation remains firmly anchored to raw pixels even at the higher levels of abstraction deep in the network. There is clearly no emergent understanding of the paddle as an actual object that can be moved. In other words, there is nothing close to a human-like comprehension of the material objects that the pixels on the screen represent or the physics that govern their movement. It’s just pixels all the way down. While some AI researchers may continue to believe that a more comprehensive understanding might eventually emerge if only there were more layers of artificial neurons, running on faster hardware and consuming still more data, I think this is very unlikely. More fundamental innovations will be required before we begin to see machines with a more human-like conception of the world.

这种一般类型的问题，即人工智能系统缺乏灵活性，无法适应输入数据中哪怕是微小的意外变化，研究人员称之为“脆弱性”。如果脆弱的人工智能应用导致仓库机器人偶尔将错误的物品装进箱子里，那么它可能不会是一个大问题。然而，在其他应用中，同样的技术缺陷可能是灾难性的。例如，这解释了为什么全自动驾驶汽车的进展未能达到一些较为乐观的早期预测。

This general type of problem, in which an AI system is inflexible and unable to adapt to even small unexpected changes in its input data, is referred to, among researchers, as “brittleness.” A brittle AI application may not be a huge problem if it results in a warehouse robot occasionally packing the wrong item into a box. In other applications, however, the same technical shortfall can be catastrophic. This explains, for example, why progress toward fully autonomous self-driving cars has not lived up to some of the more exuberant early predictions.

随着这些局限性在 2020 年末逐渐成为焦点，人们开始担心这个领域再次陷入困境，炒作周期将人们的期望推向了不切实际的水平。在科技媒体和社交媒体上，人工智能领域最可怕的短语之一——“人工智能寒冬”——再次出现。在 2020 年 1 月接受 BBC 采访时，Yoshua Bengio 表示，“人工智能的能力被某些对此感兴趣的公司夸大了。” ¹⁴

As these limitations came into focus toward the end of the decade, there was a gnawing fear that the field had once again gotten over its skis and that the hype cycle had driven expectations to unrealistic levels. In the tech media and on social media, one of the most terrifying phrases in the field of artificial intelligence—“AI winter”—was making a reappearance. In a January 2020 interview with the BBC, Yoshua Bengio said that “AI’s abilities were somewhat overhyped… by certain companies with an interest in doing so.”¹⁴

这种担忧很大一部分都影响到了处于所有炒作顶峰的行业：自动驾驶汽车。很明显，尽管本世纪初人们对自动驾驶汽车做出了乐观的预测，但能够在各种条件下行驶的真正无人驾驶汽车仍然离现实还很远。Waymo、Uber 和 Tesla 等公司已将自动驾驶汽车投入公共道路，但除了一些非常有限的实验外，一直都有人类司机——事实证明，人类司机不得不经常控制汽车。即使有司机监督汽车的运行，一些致命事故也损害了该行业的声誉。在 2018 年广为流传的一篇题为“人工智能寒冬即将来临”的博客文章中，机器学习研究员 Filip Piekniewski 指出，加利福尼亚州要求的记录显示，一辆正在测试的汽车“真的无法行驶十英里”，除非系统脱离，否则需要人类司机接管控制^。15

A large share of this concern came to bear on the industry that was, as we saw in Chapter 3, at the absolute summit of all the accumulated hype: self-driving cars. It was becoming clear that, despite optimistic predictions early in the decade, truly driverless vehicles, capable of navigating in a wide range of conditions, were still not close to reality. Companies like Waymo, Uber and Tesla had put autonomous vehicles on public roads, but outside a few very constrained experiments, there was always a human driver—who, it turned out, had to take control of the car all too often. Even with a driver in place to oversee the car’s operation, a number of fatal accidents had tarnished the industry’s reputation. In a widely shared 2018 blog post entitled “AI Winter Is Well on Its Way,” machine learning researcher Filip Piekniewski pointed out that records required by the State of California showed that one car being tested “literally could not drive ten miles” without a system disengagement that required the human driver to take control.¹⁵

我个人认为，如果另一个人工智能寒冬真的来临，那它很可能是一个温和的寒冬。尽管对进展放缓的担忧是有道理的，但过去几年来，人工智能已经深深融入了大型科技公司的基础设施和商业模式。这些公司在计算资源和人工智能人才方面的巨额投资已经获得了可观的回报，他们现在认为人工智能对于他们在市场上的竞争力至关重要。同样，几乎每一家科技初创公司现在都在某种程度上投资人工智能，其他行业的大大小小的公司也开始部署这项技术。这种成功融入商业领域的现象比以往人工智能寒冬时期的任何情况都重要得多，因此，该领域受益于整个企业界的一大批支持者，并且具有总体势头，将起到缓和任何衰退的作用。

My own view is that if another AI winter indeed looms, it’s likely to be a mild one. Though the concerns about slowing progress are well founded, it remains true that over the past few years AI has been deeply integrated into the infrastructure and business models of the largest technology companies. These companies have seen significant returns on their massive investments in computing resources and AI talent, and they now view artificial intelligence as absolutely critical to their ability to compete in the marketplace. Likewise, nearly every technology startup is now, to some degree, investing in AI, and companies large and small in other industries are beginning to deploy the technology. This successful integration into the commercial sphere is vastly more significant than anything that existed in prior AI winters, and as a result the field benefits from an army of advocates throughout the corporate world and has a general momentum that will act to moderate any downturn.

可扩展性作为进步的主要驱动力的衰落也可能有好的一面。当人们普遍认为只需投入更多计算资源即可解决问题时，人们投资于真正创新的更困难的工作的动力就会大大减少。例如，摩尔定律就是这种情况。当人们几乎完全相信计算机速度大约每两年翻一番时，半导体行业倾向于专注于生产英特尔和摩托罗拉等公司相同微处理器设计的更快版本。近年来，原始计算机速度的加速变得不那么可靠，随着芯片上电路的尺寸缩小到接近原子尺寸，我们对摩尔定律的传统定义正在接近其终点。这迫使工程师们进行更多“开箱即用”的思考，从而产生了创新，例如为大规模并行计算设计的软件和全新的芯片架构——其中许多针对深度神经网络所需的复杂计算进行了优化。我认为，我们可以期待在深度学习和更广泛的人工智能领域发生同样的创意爆炸，因为简单地扩展到更大的神经网络已经不再是一条可行的进步之路。

There’s also a sense in which the fall of scalability as the primary driver of progress may have a bright side. When there is a widespread belief that simply throwing more computing resources at a problem will produce important advances, there is significantly less incentive to invest in the much more difficult work of true innovation. This was arguably the case, for example, with Moore’s Law. When there was near absolute confidence that computer speeds would double roughly every two years, the semiconductor industry tended to focus on cranking out ever faster versions of the same microprocessor designs from companies like Intel and Motorola. In recent years, the acceleration in raw computer speeds has become less reliable, and our traditional definition of Moore’s Law is approaching its end game as the dimensions of the circuits imprinted on chips shrink to nearly atomic size. This has forced engineers to engage in more “out of the box” thinking, resulting in innovations such as software designed for massively parallel computing and entirely new chip architectures—many of which are optimized for the complex calculations required by deep neural networks. I think we can expect the same sort of idea explosion to happen in deep learning, and artificial intelligence more broadly, as the crutch of simply scaling to larger neural networks becomes a less viable path to progress.

追求更通用的机器智能

THE QUEST FOR MORE GENERAL MACHINE INTELLIGENCE

要克服深度学习系统目前的局限性，就需要创新，让机器智能越来越接近人脑的能力。这条路上有许多重大障碍，但它的最终目标是实现人工智能的圣杯：一种能够在人类水平甚至更高水平上交流、推理和构思新想法的机器。研究人员通常将其称为“通用人工智能”，简称 AGI。目前现实世界中还没有接近 AGI 的东西，但在科幻小说中有很多例子——包括《2001：太空漫游》中的 HAL、 《星际迷航》中的企业号主计算机和 Mr. Data ，当然还有《终结者》和《黑客帝国》中描绘的真正反乌托邦的技术。人们可以有力地论证说，具有超人能力的通用机器智能的发展将是人类历史上最重要的创新：这种技术将成为终极智力工具，大大加快无数领域的进步速度。在人工智能专家中，对于实现 AGI 需要多长时间，意见分歧很大。一些研究人员非常乐观，认为这一突破可能在五到十年内发生。其他人则更加谨慎，认为可能需要一百年甚至更长时间。

Overcoming the current limitations of deep learning systems will require innovations that bring machine intelligence inexorably closer to the capabilities of the human brain. There are many significant obstacles along this path, but it culminates in what has always been the Holy Grail of artificial intelligence: a machine that can communicate, reason and conceive new ideas at the level of a human being or beyond. Researchers often refer to this as “artificial general intelligence,” or AGI. Nothing close to AGI currently exists in the real world, but there are many examples from science fiction—including HAL from 2001: A Space Odyssey, the Enterprise main computer and Mr. Data from Star Trek and, of course, the truly dystopian technologies portrayed in The Terminator and The Matrix. One can make a strong argument that the development of general machine intelligence with superhuman capability would be the most consequential innovation in the history of humanity: such a technology would become the ultimate intellectual tool, dramatically accelerating the rate of progress in countless areas. Among AI experts, opinions vary widely as to just how long it might take to achieve AGI. A few researchers are extremely optimistic, suggesting this breakthrough could occur within five to ten years. Others are far more cautious and believe that it might take a hundred years or more.

在可预见的未来，大多数研究的重点不是人类水平的人工智能的实际成就，而是通往这一目标的旅程，以及成功克服沿途障碍所需的众多重要创新。打造一台真正的思考机器的探索不仅仅是一个推测性的科学项目；相反，它代表了一种路线图，通往构建能够克服当前限制并展现新功能的人工智能系统的道路。沿着这条道路前进几乎肯定会产生大量具有巨大商业和科学价值的实际应用。

For the foreseeable future, most research is focused not so much on the actual achievement of human-level AI, but rather on the journey toward it and the numerous important innovations that will be required to successfully navigate the obstacles along the way. The quest to build a true thinking machine is not just a speculative science project; rather, it represents a kind of road map toward building AI systems that overcome current limitations and exhibit new capabilities. Progress along that path is nearly certain to spawn a wealth of practical applications with enormous commercial and scientific value.

这种将实用的短期创新与对真正人类水平的机器智能的更宏伟追求相结合的做法，体现在谷歌各个 AI 团队的研究理念中。谷歌人工智能总总监 Jeff Dean 告诉我，虽然谷歌 2014 年收购的独立公司 DeepMind 专门面向通用机器智能，并制定了“结构化计划”来解决特定问题，希望最终实现 AGI，但谷歌的其他研究小组则采取了“更有机”的方法，专注于做“我们知道很重要但目前无法做到的事情，一旦解决了这些问题，我们就会弄清楚下一组我们想要解决的问题是什么”。他说，谷歌的所有 AI 研究小组都在“共同努力，试图构建真正智能、灵活的 AI 系统”。16只有时间才能告诉我们^，自上而下的计划方法和循序渐进的探索过程哪种方法更成功，但这两条路径都很可能产生重要的新想法并立即应用。

This coupling of practical near-term innovation with the far more aspirational quest for true human-level machine intelligence is demonstrated by the research philosophies of the various teams working on AI at Google. Jeff Dean, Google’s overall director of artificial intelligence, told me that while DeepMind, the independent company Google acquired in 2014, is specifically oriented toward general machine intelligence with a “structured plan” to solve specific issues in the hope of eventually achieving AGI, other research groups at Google take a “more organic” approach, with a focus on doing things “that we know are important but that we can’t do yet, and once we solve those, then we figure out what is the next set of problems that we want to solve.” All the AI research groups at Google, he says, are “working together on trying to build really intelligent, flexible AI systems.”¹⁶ Only time will tell whether a top-down planned approach or a process of step-by-step exploration will be more successful, but both paths will in all likelihood generate important new ideas with immediate applications.

这些研究道路上的进展由拥有不同研究理念和应对未来挑战不同策略的团队领导。他们的共同点是，他们的最终目标都是以至少到目前为止人类认知所独有的能力为模型。

Progress along these paths is being led by teams with varying research philosophies and many different strategies for confronting the challenges that lie ahead. What they all have in common is that their ultimate objectives are modeled on capabilities that, at least so far, are exclusive to human cognition.

一个重要的方法是直接观察人脑的内部运作来寻找灵感。这些研究人员认为人工智能应该直接受到神经科学的影响。该领域的领导者是 DeepMind。该公司的创始人兼首席执行官 Demis Hassabis——对于人工智能研究人员来说很不寻常——接受了神经科学而非计算方面的研究生培训，并拥有伦敦大学学院该领域的博士学位。Hassabis 告诉我，DeepMind 最大的研究小组由神经科学家组成，他们专注于寻找将脑科学的最新见解应用于人工智能的方法^。17

One important approach is to look directly to the inner workings of the human brain for inspiration. These researchers believe that artificial intelligence should be directly informed by neuroscience. The leader in this area is DeepMind. The company’s founder and CEO, Demis Hassabis—unusually for an AI researcher—received his graduate training in neuroscience, rather than computing, and holds a PhD in the field from University College, London. Hassabis told me that the single largest research group at DeepMind consists of neuroscientists who are focused on finding ways to apply the latest insights from brain science to artificial intelligence.¹⁷

他们的目标不是复制大脑运作方式的任何细节，而是从大脑运作的基本原理中汲取灵感。人工智能专家经常用动力飞行的实现和随后现代飞机设计的发展来类比，以解释这种方法。虽然飞机显然是受鸟类启发而来的，但它们当然不会拍打翅膀或试图直接模仿鸟类飞行。相反，一旦工程师掌握了空气动力学科学，就有可能制造出按照鸟类飞行基本原理运行的机器，但在大多数方面，这些机器的能力远超其生物同类。哈萨比斯和 DeepMind 团队认为，可能存在一种“智能空气动力学”——一种人类和潜在的机器智能的基础理论。

Their objective is not to replicate the way the brain works in any detailed sense, but rather to be inspired by the basic principles that underlie its operation. AI experts often explain this approach by using an analogy to the achievement of powered flight and the subsequent development of modern aircraft designs. While airplanes are clearly inspired by birds, they do not, of course, flap their wings or otherwise attempt to directly mimic avian flight. Rather, once engineers understood the science of aerodynamics, it became possible to build machines that operate according to the same fundamental principles that allow birds to fly, but which are in most ways far more capable than their biological counterparts. Hassabis and the team at DeepMind believe that there might be a kind of “aerodynamics of intelligence”—a foundational theory that underlies human and, potentially, machine intelligence.

2018 年 5 月，DeepMind 的跨学科团队发表了一项研究成果，并提供了令人信服的证据，证明这样一套通用原则确实存在。四年前，三位神经科学家约翰·奥基夫 (John O'Keefe)、梅-布里特·莫泽 (May-Britt Moser) 和爱德华·莫泽 (Edvard Moser) 因发现一种特殊的神经元而获得了诺贝尔生理学或医学奖，这种神经元可以让动物实现空间导航。当动物探索周围环境时，这些神经元被称为网格细胞，它们会在大脑内以规则的六边形模式激发。网格细胞被认为是一种“内部 GPS”，即一种映射系统的神经表征，可以让动物在复杂且不可预测的环境中寻找方向时保持方向。

DeepMind’s cross-disciplinary team delivered some compelling evidence that such a general set of principles might indeed exist when the company published research in May 2018. Four years earlier, the Nobel Prize in Physiology or Medicine had been awarded to three neuroscientists, John O’Keefe, May-Britt Moser and Edvard Moser, for their discovery of a special type of neuron that enables spatial navigation in animals. These neurons, called grid cells, fire in a regular hexagonal pattern within the brain as the animal explores its environment. Grid cells are thought to make up a kind of “internal GPS,” a neural representation of a mapping system that allows animals to remain oriented as they find their way through complex and unpredictable environments.

DeepMind 进行了一项计算实验，该公司的研究人员用模拟动物在黑暗中觅食时可能依赖的基于运动的信息的数据训练了一个强大的神经网络。值得注意的是，研究人员发现网格细胞状结构“自发地出现在网络中——与觅食哺乳动物中观察到的神经活动模式惊人地一致。” ¹⁸换句话说，相同的基本导航结构似乎自然出现在两个完全不同的基质中，一个是生物的，另一个是数字的。哈萨比斯告诉我，他认为这是该公司最重要的突破之一，这项研究可能表明，利用网格细胞的内部系统可能是在任何系统中表示导航信息最具计算效率的方式，无论其实施细节如何。¹⁹ DeepMind 在《自然》杂志上发表了描述这项研究的科学论文，²⁰在神经科学领域引起了广泛共鸣，这样的见解表明，该公司的跨学科方法可能会变成一条双向的道路，人工智能研究不仅借鉴大脑的经验教训，而且还有助于对大脑的理解。

DeepMind conducted a computational experiment in which the company’s researchers trained a powerful neural network on data that simulated the kind of movement-based information that an animal foraging in the dark might rely on. Remarkably, the researchers found that grid cell–like structures “spontaneously emerged within the network—providing a striking convergence with the neural activity patterns observed in foraging mammals.”¹⁸ In other words, the same fundamental navigation structure appears to arise naturally in two entirely different substrates, one biological and the other digital. Hassabis told me he considers this to be one of the company’s most important breakthroughs and that the research may indicate that an internal system utilizing grid cells may simply be the most computationally efficient way to represent navigation information in any system, regardless of the details of its implementation.¹⁹ DeepMind’s scientific paper describing the research, published in the journal Nature,²⁰ resonated widely within the field of neuroscience, and insights like this suggest the company’s interdisciplinary approach will likely turn out to be a two-way street, with AI research not only drawing upon lessons from the brain but also contributing to its understanding.

2020 年初，DeepMind 利用其在强化学习方面的专业知识，探索大脑中多巴胺神经元的运作，再次为神经科学做出了重要贡献。21自 20 世纪⁹⁰年代以来，神经科学家已经了解到，这些特殊的神经元会预测动物采取特定行动时可能产生的奖励。如果事实证明实际获得的奖励大于预期，那么就会释放出相对更多的多巴胺。如果结果不尽如人意，那么就会产生较少的这种“感觉良好”的化学物质。计算强化学习的传统工作方式大致相同；算法进行预测，然后根据预测结果和实际结果之间的差异调整奖励。

DeepMind once again made an important contribution to neuroscience when the company leveraged its expertise in reinforcement learning in early 2020 by exploring the operation of dopamine neurons in the brain.²¹ Since the 1990s, neuroscientists have understood that these special neurons make a prediction about the likely reward that will result when an animal takes a specific action. If it turns out that the reward actually achieved is greater than expected, then relatively more dopamine is released. If the result underperforms, then less of this “feel good” chemical is generated. Computational reinforcement learning traditionally works in much the same way; the algorithm makes a prediction and then adjusts the reward based on the difference between the predicted and actual results.

DeepMind 的研究人员通过生成预测分布（而不是单一的平均预测）并相应地调整奖励，极大地改进了强化学习算法。该公司随后与哈佛大学的一组研究人员合作，研究大脑中是否会发生同样的事情。他们能够证明，小鼠大脑实际上采用了类似的预测分布，一些多巴胺神经元相对更悲观，而另一些多巴胺神经元对潜在奖励则更乐观。换句话说，该公司再次证明了同样的基本机制在数字算法和生物大脑中实现了并行结果。

Researchers at DeepMind were able to greatly improve a reinforcement learning algorithm by generating a distribution of predictions, rather than a single average prediction, and then adjusting the rewards accordingly. The company then teamed with a group of researchers at Harvard to see if the same kind of thing might be happening within the brain. They were able to show that mouse brains actually employ a similar distribution of predictions, with some dopamine neurons being relatively more pessimistic and others more optimistic about the potential reward. In other words, the company had once again demonstrated the same fundamental mechanism achieving parallel results in both a digital algorithm and the biological brain.

这类研究反映了哈萨比斯及其团队对强化学习的信心，以及他们坚信强化学习是迈向更通用人工智能的任何尝试的关键组成部分。在这方面，他们有点与众不同。例如，Facebook 的 Yann LeCun 表示，他认为强化学习起着相对较小的作用。他在演讲中经常说，如果智能是一块黑森林蛋糕，那么强化学习只不过是蛋糕顶部的樱桃。22 ^DeepMind团队认为强化学习更为重要，并且它可能为实现 AGI 提供了一条可行的途径。

Research of this type reflects the confidence that Hassabis and his team have in reinforcement learning and their belief that it is a critical component of any attempt to progress toward more general artificial intelligence. In this, they are something of an outlier. Facebook’s Yann LeCun, for example, has stated that he believes reinforcement learning plays a relatively minor role. In his presentations, he often says that if intelligence were a black forest cake, then reinforcement learning would amount to only the cherry on top.²² The team at DeepMind believes it is far more central—and that it possibly provides a viable path to achieving AGI.

我们通常将强化学习描述为一种奖励驱动算法，用于优化某些外部宏观过程——例如，学习下围棋或弄清楚如何驾驶模拟汽车。然而，哈萨比斯指出，强化学习在大脑内部也发挥着关键作用，它可能对智力的出现至关重要。可以想象，强化学习可能是驱使大脑产生好奇心、学习和推理的主要机制。例如，想象一下，大脑的固有目标只是探索，然后整理动物在环境中移动时不断轰击的大量原始数据。哈萨比斯说，“我们知道看到新奇的事物会释放大脑中的多巴胺”，如果大脑的构造使得“寻找信息和结构本身就是一种奖励，那么这是一种非常有用的动机。” ²³换句话说，驱动我们不断了解周围世界的引擎可能是一种与多巴胺产生有关的强化学习算法。

We generally describe reinforcement learning in terms of a reward-driven algorithm to optimize some external macro process—for example, learning to play the game of Go or figuring out how to drive a simulated car. However, Hassabis points out that reinforcement learning also plays a critical role internal to the brain, and that it may be essential to the emergence of intelligence. It’s conceivable that reinforcement learning might be the primary mechanism that drives the brain toward curiosity, learning and reason. Imagine, for example, that the brain’s inherent objective is simply to explore and then bring order to the torrent of raw data constantly bombarding an animal as it moves through its environment. Hassabis says that “we know seeing novel things releases dopamine in the brain” and if the brain is wired so that “finding information and structure [is] rewarding in itself, then that’s a highly useful motivation.”²³ In other words, the engine powering our continuous drive to understand the world around us could be a reinforcement learning algorithm linked to the production of dopamine.

人工智能初创公司 Elemental Cognition 的首席执行官兼创始人 David Ferrucci 正在寻求一种完全不同的方法来构建更通用的机器智能。Ferrucci 最出名的是领导了 IBM Watson 的团队，该系统在 2011 年击败了 Ken Jennings 和其他Jeopardy!顶级参赛者。在 Watson 获胜后，Ferrucci 离开 IBM 并加入了华尔街对冲基金 Bridgewater and Associates，据报道，他在那里致力于利用人工智能来理解宏观经济，并帮助将 Bridgewater 创始人 Ray Dalio 的管理和投资理念融入到整个公司部署的算法中。

An entirely different approach to building more general machine intelligence is being pursued by David Ferrucci, the CEO and founder of the AI startup company Elemental Cognition. Ferrucci is best known for leading the team that created IBM’s Watson, the system that defeated Ken Jennings and other top Jeopardy! contestants in 2011. After Watson’s triumph, Ferrucci left IBM and joined the Wall Street hedge fund Bridgewater and Associates, where he reportedly worked on using artificial intelligence to make sense of the macroeconomy and helped build Bridgewater founder Ray Dalio’s management and investment philosophies into algorithms deployed throughout the firm.

Ferrucci 现在一边担任 Bridgewater 应用 AI 总监，一边管理 Element Cognition，该公司的初始风险投资资金来自 Bridgewater。24 Ferrucci 告诉我，Elemental Cognition 专注于“真实语言理解”。该公司正在构建能够自主阅读文本然后与人类进行互动对话的算法，以增强系统对材料的理解并解释任何结论。Ferrucci 继续说道^：

Ferrucci now splits his time between his position as director of applied AI at Bridgewater and running Elemental Cognition, which received initial venture funding from the hedge fund.²⁴ Ferrucci told me that Elemental Cognition is focused on “real language understanding.” The company is building algorithms that can autonomously read text and then engage in interactive dialog with humans in order to enhance the system’s understanding of the material and also explain any conclusions. Ferrucci goes on to say:

我们希望超越语言的表面结构，超越词频模式，了解其深层含义。从中，我们希望能够构建人类将创建并用于推理和交流的内部逻辑模型。我们希望确保系统能够产生兼容的智能。兼容的智能可以通过人类互动、语言、对话和其他相关经验自主学习和完善其理解^。25

We want to look beyond the surface structure of language, beyond the patterns that appear in word frequencies, and get at the underlying meaning. From that, we want to be able to build the internal logical models that humans would create and use to reason and communicate. We want to ensure a system that produces a compatible intelligence. That compatible intelligence can autonomously learn and refine its understanding through human interaction, language, dialog, and other related experiences.²⁵

这是一个非常雄心勃勃的目标，在我看来，它已经非常接近人类水平的智能。现有的处理自然语言的人工智能系统也存在类似的限制，就像 DeepMind 的 Atari DQN 在游戏手柄向上移动几个像素时遇到的限制一样。正如 DQN 无法理解屏幕上的像素代表可以移动的物理对象一样，当前的语言系统也无法真正理解它们处理的单词的含义。这正是 Elemental Cognition 正在应对的挑战。

This is an extraordinarily ambitious goal that, to me, sounds very close to human-level intelligence. Existing AI systems that process natural language suffer from a similar limitation to the one we saw with DeepMind’s Atari-playing DQN when the game paddle was shifted a few pixels higher. Just as DQN has no understanding that the pixels on the screen represent a physical object that can be moved, current language systems have no real comprehension of what the words they process mean. This is the challenge that Elemental Cognition is taking on.

Ferrucci 显然认为，解决语言理解问题是实现更通用智能的最清晰途径。Ferrucci 并不像 DeepMind 团队那样深入研究大脑的生理学，而是认为可以直接设计一个系统，使其在语言理解和运用逻辑和推理的能力上接近人类水平。在人工智能研究人员中，他与众不同之处在于，他认为通用智能的基本构建模块已经到位，或者用他的话说，“我不认为，其他人可能认为我们不知道如何做到这一点，我们正在等待一些巨大的突破。我不认为情况如此；我认为我们确实知道如何做到这一点，我们只需要证明这一点。” ²⁶

Ferrucci clearly believes that solving the language understanding problem represents the clearest path to more general intelligence. Rather than delving into the physiology of the brain in the way that DeepMind’s team is attempting, Ferrucci argues that it is possible to directly engineer a system that can approach human level in its comprehension of language and its ability to employ logic and reason. He is unusual among AI researchers in that he feels the basic building blocks for general intelligence are already in place, or as he put it, “I don’t think, as other people might, that we don’t know how to do it and we’re waiting for some enormous breakthrough. I don’t think that’s the case; I think we do know how to do it, we just need to prove that.”²⁶

他还对在不久的将来实现这一目标的前景非常乐观。在 2018 年的一部纪录片中，他说：“三到五年后，我们将拥有一个能够自主学习理解和建立理解的计算机系统，这与人类思维的工作方式并无二致。” ²⁷当我追问他这一预测时，他有所退缩，承认三到五年可能确实是乐观的。然而，他说他仍然会“认为这是我们可以在未来十年左右看到的事情。这不是 50 年或 100 年的等待。” ²⁸

He’s also very optimistic about the prospects for achieving this goal in the relatively near future. In a 2018 documentary film, he said, “In three to five years, we’ll have a computer system that can autonomously learn to understand and how to build understanding, not unlike the way a human mind works.”²⁷ When I pressed him on this prediction, he backed off somewhat, acknowledging that that three to five years might indeed be optimistic. However, he said he would still “argue that it’s something that we could see within the next decade or so. It’s not going to be a 50- or a 100-year wait.”²⁸

为了实现这一目标，Elemental Cognition 团队正在构建一种混合系统，该系统包括深度神经网络以及其他机器学习方法，以及使用传统编程技术构建的软件模块来处理逻辑和推理。正如我们将看到的，关于这种混合方法与完全基于神经网络的策略相比的有效性的争论正在成为人工智能领域面临的最重要问题之一。

To accomplish this goal, the team at Elemental Cognition is building a kind of hybrid system that includes deep neural networks as well as other machine learning approaches in combination with software modules built using traditional programming techniques to handle logic and reasoning. As we will see, the debate over the efficacy of such a hybrid approach as opposed to a strategy based entirely on neural networks is emerging as one of the most important questions confronting the field of AI.

雷·库兹韦尔 (Ray Kurzweil) 现为谷歌工程总监，他同样在追求通用智能，其方向主要是理解语言。库兹韦尔因其 2005 年出版的《奇点临近》一书而闻名，²⁹这本书使他成为“奇点”理念最著名的倡导者。库兹韦尔和他的众多追随者认为，奇点很可能是由超人机器智能的出现所带来的，有朝一日，奇点将标志着人类历史弧线的突然上升——一个技术加速变得如此极端的转折点，以至于彻底、甚至令人难以理解地改变人类生活和文明的各个方面。

Ray Kurzweil, now a director of engineering at Google, is likewise pursuing general intelligence along a path that is heavily oriented toward understanding language. Kurzweil is famous for his 2005 book The Singularity Is Near,²⁹ which established him as the most prominent evangelist for the idea of the “Singularity.” Kurzweil and his many followers believe that the Singularity, likely brought on by the advent of superhuman machine intelligence, will someday mark an abrupt upward bend in humanity’s historical arc—an inflection point when technological acceleration will become so extreme as to completely, and perhaps incomprehensibly, transform every aspect of human life and civilization.

2012 年，库兹韦尔出版了另一本书，名为《如何创造心智》，在书中，他勾勒出了人类认知的概念模型。30^库兹韦尔认为，大脑由大约 3 亿个分层模块驱动，每个模块“都能识别连续模式并接受一定程度的变化”。31^库兹韦尔认为，这种模块化方法最终将产生一个系统，它能从比目前依赖监督学习或强化学习技术的深度学习系统少得多的数据中进行学习。当库兹韦尔找到谷歌的拉里·佩奇，为将这些想法付诸实践的企业寻求资金时，佩奇却说服他加入谷歌，利用公司巨大的计算资源来实现他的愿景。

In 2012, Kurzweil published another book entitled How to Create a Mind, in which he sketched out a conceptual model for human cognition.³⁰ According to Kurzweil, the brain is powered by around 300 million hierarchical modules, each of which “can recognize a sequential pattern and accept a certain amount of variability.”³¹ Kurzweil believes that this modular approach will ultimately result in a system that can learn from far less data than is the case with current deep learning systems that rely on supervised or reinforcement learning techniques. When Kurzweil approached Google’s Larry Page to seek funding for a venture to put these ideas into practice, Page instead convinced him to come to Google and pursue his vision by taking advantage of the company’s enormous computing resources.

几十年来，库兹韦尔一直预测，通用人工智能将在 2029 年左右实现，并且至今仍坚信这一点。与许多人工智能研究人员不同，他仍然相信图灵测试是衡量人类水平智能的有效方法。图灵测试由艾伦·图灵在其 1950 年的论文中提出，本质上相当于一个聊天会话，其中评委试图确定对话者是人类还是机器。如果评委或评委小组无法区分计算机和人类，则认为计算机通过了图灵测试。许多专家不认为图灵测试是衡量人类水平机器智能的有效方法，部分原因是它已被证明容易受到花招的影响。例如，2014 年，在英国雷丁大学举行的一场比赛中，一个模仿 13 岁乌克兰男孩的聊天机器人成功欺骗评委，让他们宣布一种算法首次通过了图灵测试。这次对话仅仅持续了五分钟，但人工智能领域几乎没有人认真对待这一说法。

Kurzweil has predicted for decades, and still believes, that AGI will be achieved sometime around the year 2029. Unlike many AI researchers, he continues to have faith in the Turing test as an effective measure of human-level intelligence. Conceived by Alan Turing in his 1950 paper, the test essentially amounts to a chat session in which a judge attempts to determine if the conversers are human or machine. If the judge, or perhaps a panel of judges, cannot distinguish the computer from a human, then the computer is said to pass the Turing test. Many experts are dismissive of the Turing test as an effective measure of human-level machine intelligence, in part because it has proven to be susceptible to gimmicks. In 2014, for example, in a contest held at the University of Reading in the United Kingdom, a chatbot that emulated a thirteen-year-old Ukrainian boy managed to fool the judges into declaring that an algorithm had, for the first time, passed the Turing test. The conversation had lasted a mere five minutes, and virtually no one in the field of artificial intelligence took the claim seriously.

尽管如此，库兹韦尔还是相信，更严格的测试版本确实可以有力地表明真正的机器智能。2002 年，库兹韦尔与软件企业家米奇·卡普尔正式打赌 20,000 美元。该赌注指定了一套复杂的规则，包括三名评委和四名参赛者：人工智能聊天机器人和三名人类陪衬。32^只有在 2029 年底之前，大多数评委在与每位参赛者进行两小时的一对一对话后认为人工智能系统是人类时，赌注才会对库兹韦尔有利。在我看来，通过这样的测试将是一个强有力的迹象，表明人类级别的人工智能已经到来。

Kurzweil nonetheless believes that a much more robust version of the test would indeed be a powerful indicator of true machine intelligence. In 2002, Kurzweil entered into a formal $20,000 bet with the software entrepreneur Mitch Kapor. The bet specifies a complex set of rules that include a three-judge panel and four contestants: the AI-powered chatbot along with three human foils.³² The bet will be decided in Kurzweil’s favor only if, by the end of the year 2029, a majority of the judges believe the AI system to be human after engaging in a two-hour, one-on-one conversation with each of the contestants. It does seem to me that passing such a test would be a strong indication that human-level AI has arrived.

尽管库兹韦尔的发明生涯十分杰出，但现在人们往往认为他主要是一位未来学家，他有一套关于长期技术加速的合理理论，但同时也有一些看似古怪的想法，有些人甚至认为这些进步可能带来什么。有报道称，库兹韦尔每天服用一百多片补充剂，希望延长自己的寿命。33^事实上，他相信自己已经达到了“长寿逃逸速度”——换句话说，他希望能够反复活得足够长，以便利用下一次延长生命的医疗创新。34^无限期地这样做，同时避免与公交车相撞，你将获得永生。库兹韦尔告诉我，大约十年内，我们其他人应该可以实现这样的计划。他认为，将先进的人工智能应用于高保真生物化学模拟是这一进步的关键驱动力。 “如果我们能够模拟生物学，这并非不可能，那么我们可以在几小时而不是几年内进行临床试验，我们就可以生成我们自己的数据，就像我们对自动驾驶汽车、棋盘游戏或数学所做的那样，” ³⁵他告诉我。

Despite his distinguished career as an inventor, Kurzweil now tends to be perceived primarily as a futurist with a reasonably well-formulated theory about long-term technological acceleration, but also some seemingly outlandish—some might even say kooky—ideas about where all this progress is, in his view, likely to lead. By one account, Kurzweil takes one hundred or more supplement pills each day in the hope of prolonging his life.³³ Indeed, he believes he has already achieved “longevity escape velocity”—or in other words, he expects to repeatedly live long enough to take advantage of the next life-prolonging medical innovation.³⁴ Do this indefinitely, while avoiding a run-in with the proverbial bus, and you will have achieved immortality. Kurzweil told me that within about ten years, such a plan should be accessible to the rest of us. He sees the application of advanced artificial intelligence to a high-fidelity simulation of biochemistry as a critical driver of this progress. “If we could simulate biology, and it’s not impossible, then we could do clinical trials in hours rather than years, and we could generate our own data just like we’re doing with self-driving cars or board games or math,”³⁵ he told me.

诸如此类的想法，尤其是他对自己永生的坚信，让库兹韦尔遭到了不少嘲笑，许多其他人工智能研究人员对他实现通用智能的分层方案不屑一顾。然而，我与库兹韦尔交谈的主要收获之一是，谈到他在谷歌从事的人工智能工作，他似乎非常有根基。自 2012 年加入谷歌以来，他一直领导着一个团队，致力于将他的大脑分层理论与深度学习的最新进展相结合，以打造具有高级语言能力的系统。他的努力的一个早期成果是“智能回复”功能，可以在 Gmail 中提供现成的答案。虽然这无疑与人类级别的人工智能相去甚远，但库兹韦尔仍然对他的策略充满信心，他告诉我，“人类使用这种分层方法”，最终它将“足以实现通用人工智能” ^。36

Ideas like these, and perhaps in particular his earnest faith in the likelihood of his own immortality, leave Kurzweil open to a fair amount of ridicule, and many other AI researchers have a dismissive view of his hierarchical scheme for achieving general intelligence. However, one of my main takeaways from my conversation with Kurzweil was that, when it comes to the work he is doing on AI at Google, he seems extraordinary well grounded. Since he joined the company in 2012, he has been leading a team focused on merging his hierarchical theory of the brain with the latest advances in deep learning in order to produce systems with advanced language capability. One early result of his effort is the “Smart Reply” feature that can give ready answers in Gmail. While this is admittedly a far cry from human-level AI, Kurzweil remains confident in his strategy, telling me that “humans use this hierarchical approach” and that ultimately it will be “sufficient for AGI.”³⁶

另一条通往通用人工智能的道路正在由 OpenAI 开辟。OpenAI 是一家总部位于旧金山的研究机构，成立于 2015 年，资金来自伊隆·马斯克、彼得·泰尔和领英联合创始人里德·霍夫曼等人。OpenAI 最初是作为一个非营利组织成立的，其使命是开展安全且合乎道德的 AGI 探索。该组织成立的部分原因是伊隆·马斯克对超人机器智能有朝一日可能对人类构成真正威胁的深切担忧。从一开始，OpenAI 就吸引了该领域的一些顶尖研究人员，其中包括 Ilya Sutskever，他是杰夫·辛顿多伦多大学实验室团队成员，该团队构建了在 2012 年 ImageNet 竞赛中获胜的神经网络。

Yet another path toward artificial general intelligence is being forged by OpenAI, a San Francisco–based research organization that was founded in 2015 with financial backing from, among others, Elon Musk, Peter Thiel and Linked-in co-founder Reid Hoffman. OpenAI was initially set up as a nonprofit entity with a mission to undertake a safe and ethical quest for AGI. The organization was conceived partly in response to Elon Musk’s deep concern about the potential for superhuman machine intelligence to someday pose a genuine threat to humanity. From the onset, OpenAI has attracted some of the field’s top researchers, including Ilya Sutskever, who was part of the team from Geoff Hinton’s University of Toronto Lab that built the neural network that triumphed at the 2012 ImageNet competition.

2019 年，当时负责硅谷最知名的初创企业孵化器 Y-Combinator 的 Sam Altman 出任首席执行官，并进行了复杂的法律重组，最终将原来的非营利实体并入了营利性公司。这样做是为了吸引足够的私营部门投资，以便 OpenAI 能够为计算资源的大规模投资提供资金，并争夺日益稀缺的人工智能人才。这一举措很快获得了回报：2019 年 7 月，微软宣布将向新公司投资 10 亿美元。

In 2019, Sam Altman, who was then in charge of Silicon Valley’s highest profile startup incubator, Y-Combinator, became CEO and undertook a complicated legal reshuffling that resulted in a for-profit company attached to the original nonprofit entity. This was done in order to attract enough investment from the private sector so that OpenAI could fund massive investment in computational resources and compete for increasingly scarce AI talent. The maneuver paid quick dividends: in July 2019, Microsoft announced it would make a billion-dollar investment in the new company.

在 AGI 的竞争中，OpenAI 可能是谷歌 DeepMind 资金最雄厚的竞争对手，尽管在员工人数方面，OpenAI 仍比这家更成熟的公司小得多。与 DeepMind 一样，OpenAI 也使用强化学习技术训练开发了强大的深度神经网络，其研究团队已经创建了能够在电子游戏中击败人类最优秀玩家的系统，例如Dot 2。然而，OpenAI 的与众不同之处在于，它专注于构建在越来越强大的计算平台上运行的越来越大的深度神经网络。尽管该领域的其他同行警告说，可扩展性作为一种战略正在变得不可持续，但 OpenAI 仍然在这种方法上投入了大量资金。事实上，微软的数十亿美元投资将主要以这家科技巨头的云计算业务 Azure 提供的计算能力的形式实现。

In the race for AGI, OpenAI is probably the best funded competitor to Google’s DeepMind, although in terms of staffing levels, it remains far smaller than the more established company. Like DeepMind, OpenAI has developed powerful deep neural networks trained using reinforcement learning techniques, and its team of researchers has created systems capable of defeating the best human players at video games, such as Dot 2. OpenAI, however, sets itself apart with a singular focus on building larger and larger deep neural networks running on ever more powerful computational platforms. Even as others in the field warn that scalability is becoming unsustainable as a strategy, OpenAI remains deeply invested in the approach. Indeed, Microsoft’s billion-dollar investment will be delivered largely in the form of compute power provided by the tech giant’s cloud computing business, Azure.

毫无疑问，OpenAI 的“越大越好”理念已经取得了重大进展。该组织最引人注目、也是最具争议的突破之一是 2019 年 2 月展示了强大的自然语言系统 GPT-2。GPT-2 由一个“生成”神经网络组成，该网络已使用从互联网下载的大量文本进行训练。在生成系统中，深度神经网络的输出本质上被翻转，因此系统不是识别或分类数据（比如为照片配上标题），而是创建与训练数据大致相似的全新示例。生成深度学习系统是所谓深度伪造背后的技术——媒体伪造很难甚至不可能与真实内容区分开来。深度伪造是与人工智能相关的一个关键风险因素，我们将在第 8 章中讨论其影响。

OpenAI’s “bigger is better” mentality has, to be sure, produced significant progress. One of the organization’s most notable, and controversial, breakthroughs came with the demonstration of a powerful natural language system called GPT-2 in February 2019. GPT-2 consists of a “generative” neural network that has been trained on a massive trove of text downloaded from the internet. In a generative system, the output of a deep neural network is essentially flipped, so that rather than identifying or classifying data—as in coming up with captions for photographs—the system instead creates entirely new examples that are broadly similar to the data it was trained on. Generative deep learning systems are the technology behind so-called deepfakes—media fabrications that can be very difficult, or perhaps impossible, to distinguish from the real thing. Deepfakes are a critical risk factor associated with artificial intelligence, and we will discuss their implications in Chapter 8.

GPT-2 的设置是这样的：只要给出一两句话的文本提示，系统就会生成一个完整的叙述——实际上是从提示结束的地方继续讲故事。GPT-2 在人工智能研究人员中引起了轰动，尤其是媒体，因为它呈现的叙述文本在大多数情况下都非常连贯。事实上，由于连贯性太强，OpenAI 拒绝与其他人工智能研究人员分享 GPT-2 代码，就像它对之前的项目所做的那样，因为担心该系统可能被滥用。例如，OpenAI 的管理团队担心，GPT-2 系统生成的大量文本可能会让互联网充斥着虚假新闻，或者让电子商务网站充斥着虚假的产品评论。这一决定遭到了人工智能研究界的批评，因为人工智能研究界重视开放的思想交流。然后，不到一年的时间，OpenAI 改变了方向，最终发布了 GPT-2 代码——没有大张旗鼓，也没有任何恶意使用的证据。这反过来又导致人们指责该公司故意炒作其研究成果，并利用危言耸听来引起宣传。

GPT-2 was set up so that, given a text prompt of perhaps a sentence or two, the system would then generate a complete narrative—in effect picking up where the prompt leaves off and completing the story. GPT-2 caused a stir among AI researchers, and especially the media, because the narrative text it rendered was, in most cases, remarkably coherent. So coherent, in fact, that OpenAI declined to share the GPT-2 code with other AI researchers, as it had done with its previous projects, out of concern that the system could be misused. OpenAI’s management team worried, for example, that a torrent of text generated by a GPT-2 system might clog the internet with fake news or perhaps overwhelm e-commerce sites with fabricated product reviews. This decision resulted in criticism from the AI research community, which values an open exchange of ideas. Then, after less than a year, OpenAI reversed course and released the GPT-2 code after all—with little fanfare and no evidence of any malicious use. This, in turn, led to accusations that the company was intentionally hyping its research and using alarmism to generate publicity.

虽然 GPT-2 生成的散文在语法和结构方面总体上还算合格，但即使粗略阅读也能清楚地发现，所讲述的故事存在一些真正的问题。以下是 OpenAI 网站上发布的 GPT-2 叙述示例的一部分：

Though the prose generated by GPT-2 is generally competent in terms of grammar and structure, even a cursory reading makes it clear that there are some real issues with the narrative being told. Here is a portion of an example of a GPT-2 narrative published on the OpenAI website:

给出人类书写的提示：

Given the human-written prompt:

令人震惊的是，一位科学家在安第斯山脉一个偏远、从未被探索过的山谷中发现了一群独角兽。更让研究人员惊讶的是，这些独角兽竟然能说一口流利的英语。

In a shocking finding, a scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.

GPT-2 生成：

GPT-2 generated:

科学家根据这些独角的独特形状，将这些动物群命名为“奥维德独角兽”。这些长着四个角、身着银白色的独角兽此前在科学界尚属未知。

The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science.

现在，经过近两个世纪，引发这一奇怪现象的谜团终于被解开了。

Now, after almost two centuries, the mystery of what sparked this odd phenomenon is finally solved.

拉巴斯大学的进化生物学家豪尔赫·佩雷斯博士和几位同伴在探索安第斯山脉时，发现了一个小山谷，那里没有其他动物或人类。佩雷斯注意到山谷里有一个看似天然喷泉的东西，周围环绕着两座岩石山峰和银色的积雪。

Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans. Pérez noticed that the valley had what appeared to be a natural fountain, surrounded by two peaks of rock and silver snow.

佩雷斯和其他人随后继续深入山谷。“当我们到达一座山峰的顶部时，水看起来是蓝色的，上面还有一些水晶，”佩雷斯说。

Pérez and the others then ventured further into the valley. “By the time we reached the top of one peak, the water looked blue, with some crystals on top,” said Pérez.

佩雷斯和他的朋友们看到独角兽群后都惊呆了。这些生物从空中就能看到，不用移动太多就可以看到它们——它们离我们如此之近，以至于可以触摸到它们的角。

Pérez and his friends were astonished to see the unicorn herd. These creatures could be seen from the air without having to move too much to see them—they were so close they could touch their horns.

在研究这些奇异生物时，科学家们发现这些生物还会说一些相当普通的英语。佩雷斯说：“例如，我们可以看到它们有一种共同的‘语言’，类似于方言或方言法。” ³⁷

While examining these bizarre creatures the scientists discovered that the creatures also spoke some fairly regular English. Pérez stated, “We can see, for example, that they have a common ‘language,’ something like a dialect or dialectic.”³⁷

故事继续了好几段，但就在开头，我们得知发现了一种新的“四角”独角兽。后来，我们被告知独角兽讲的是“普通英语”，但它们“有一种共同的‘语言’，类似于方言或方言”。人们不禁想知道这句话到底是什么意思，“这些生物可以从空中看到，而不必移动太多就能看到它们——它们离得太近，可以触摸到它们的角。”

The story continues for several more paragraphs, but right at the top we learn that a new species of “four-horned” unicorns has been discovered. Later, we’re told that the unicorns speak “regular English,” but that they “have a common ‘language,’ something like a dialect or dialectic.” And one is left wondering what exactly to make of the sentence, “These creatures could be seen from the air without having to move too much to see them—they were so close they could touch their horns.”

所有这些都清楚地表明，尽管 OpenAI 开发的庞大系统中数百万个人工神经元中确实有某种东西在凝聚，但人们并没有真正理解它。该系统不知道独角兽是什么，也不知道“四角”品种会与这一含义相矛盾。GPT-2 面临着 Elemental Cognition 的 David Ferrucci 团队和谷歌的 Ray Kurzweil 试图解决的相同基本限制。

All this makes it pretty clear that while something is indeed coalescing within the millions of artificial neurons that make up the massive system developed by OpenAI, there is no real understanding. The system does not know what a unicorn is, or that a “four-horned” variety would contradict that meaning. GPT-2 suffers from the same fundamental limitation that David Ferrucci’s team at Elemental Cognition and Ray Kurzweil at Google are trying to address.

2020 年 5 月，OpenAI 发布了 GPT-3，这是一个功能更强大的系统。GPT-2 的神经网络包含大约 15 亿个权重，这些权重在网络训练过程中得到了优化，而 GPT-3 将这一数字增加了一百多倍，达到 1750 亿个。GPT-3 的神经网络在大约半兆字节的文本上进行了训练，这个数字非常庞大，以至于整个英文版维基百科（约 600 万篇文章）仅占总数的 0.6% 左右。OpenAI 向一组精选的人工智能研究人员和记者提供了早期使用权，并宣布计划最终将新系统转变为其首个商业产品。

In May 2020, OpenAI released GPT-3, a vastly more powerful system. While GPT-2’s neural network included about 1.5 billion weights that were optimized as the network was trained, GPT-3 increased that number more than a hundredfold to 175 billion. GPT-3’s neural network was trained on about half a terabyte of text, an amount so vast that the entire English version of Wikipedia—roughly six million articles—constitutes only about 0.6 percent of the total. OpenAI offered early access to a select group of AI researchers and journalists and announced plans to eventually turn the new system into its first commercial product.

在接下来的几周里，随着人们开始尝试 GPT-3，社交媒体对新系统的强大功能感到震惊。只要给出适当的提示，GPT-3 就可以用已故作家的风格写出令人信服的文章或诗歌。它甚至可以生成历史人物或虚构人物之间的虚假对话。一名大学生使用该系统为一个自助博客生成了所有帖子，该博客登上了排行榜榜首。38^这一切很快引发了人们的猜测，即该系统代表了人类级机器智能道路上的重大突破。

Over the next few weeks, as people began to experiment with GPT-3, social media exploded with astonishment at the power of the new system. Given the proper prompts, GPT-3 could write convincing articles or poems in the style of long-dead authors. It could even generate faux conversations between historical or fictional figures. A college student used the system to generate all the posts for a self-help blog that rose to the top of the charts.³⁸ All this quickly led to speculation that the system represented a critical breakthrough on the path to human-level machine intelligence.

然而，很快就清楚的是，许多最令人印象深刻的例子都是从多次试验中精心挑选出来的，而且 GPT-3 和它的前身一样，经常会产生连贯的废话。OpenAI 的两个 GPT 系统本质上都是强大的预测引擎。给定一个单词序列，它们擅长预测下一个单词应该是什么。GPT-3 将这种能力提升到了一个前所未有的水平，而且由于系统训练的大量文本包含真实的知识，因此该系统确实经常产生非常有用的输出。然而，它没有一致性，GPT-3 经常会产生垃圾，并且在任何对人类来说都很简单的任务上都遇到了困难。39^与其前身相比，GPT-3 当然可以写出一个关于独角兽的更引人入胜的故事。然而，它仍然不了解独角兽是什么。

It soon became clear, however, that many of the most impressive examples had been cherry-picked from multiple trials, and that GPT-3, like its predecessor, often produced coherently written nonsense. Both of OpenAI’s GPT systems are at their core powerful prediction engines. Given a sequence of words, they excel at predicting what the next word should be. GPT-3 takes this capability to an unprecedented level, and because the massive trove of text the system was trained on encapsulates real knowledge, the system does often produce very useful output. There is no consistency, however, and GPT-3 often generates rubbish and struggles with tasks that would be simple for any human.³⁹ Compared to its predecessor, GPT-3 can certainly write a far more compelling story about unicorns. It still, however, has no understanding of what a unicorn is.

如果 OpenAI 只是继续将更多的计算资源投入到这个问题中，如果他们构建更大规模的神经网络，那么真正的理解是否会出现？在我看来，这似乎不太可能，许多人工智能专家对 OpenAI 对可扩展性的持续信心持批评态度。加州大学伯克利分校计算机科学教授、世界领先的人工智能大学教科书合著者 Stuart Russell 告诉我，实现 AGI 需要“与更大的数据集或更快的机器无关”的突破^。40

If OpenAI simply continues to throw more computational resources at the problem, if they build even more massive neural networks, is it likely that true understanding will emerge? To me, that seems very unlikely, and many AI experts are very critical of OpenAI’s continued faith in scalability. Stuart Russell, a professor of computer science at the University of California, Berkley, and the co-author of the world’s leading university textbook on artificial intelligence, told me that achieving AGI will require breakthroughs that “have nothing to do with bigger datasets or faster machines.”⁴⁰

尽管如此，OpenAI 团队仍然充满信心。该公司首席科学家 Ilya Sutskever 在 2018 年的一次技术会议上发表演讲时表示：“我们回顾了过去六年该领域的进展。我们的结论是，短期内 AGI 应该被视为一种严肃的可能性。” ⁴¹几个月后，在另一场会议上，OpenAI 首席执行官 Sam Altman 表示：“我确实认为，构建 [AGI] 的秘诀在于将这些系统规模扩大得越来越大。” ⁴²这种方法尚无定论，但我猜，为了取得成功，OpenAI 需要加大其在真正创新方面的努力，而不仅仅是扩大其神经网络的规模。

Still, the OpenAI team remains confident. In a speech at a technology conference in 2018, the company’s chief scientist, Ilya Sutskever, said, “We have reviewed progress in the field over the past six years. Our conclusion is near term AGI should be taken as a serious possibility.”⁴¹ Several months later, at another conference, OpenAI CEO Sam Altman said, “I do think that much of the secret to building [AGI] is just going to be scaling these systems bigger and bigger and bigger.”⁴² The jury remains out on this approach, but my guess is that in order to achieve success, OpenAI will need to scale up its efforts at genuine innovation—rather than just the size of its neural networks.

符号人工智能的复兴与先天结构的争论

A REVIVAL OF SYMBOLIC AI AND THE DEBATE OVER INNATE STRUCTURE

在研究人员努力应对未来挑战的同时，符号人工智能阵营所倡导的理念正在经历某种复兴。几乎每个人都承认，如果人工智能要向前发展，就必须解决符号学派试图解决但大多以失败告终的问题。除了相对较少的深度学习纯粹主义者（其中许多人似乎与 OpenAI 有关），研究人员几乎没有信心，仅仅扩展现有的神经算法以利用更快的硬件和更多的数据就足以产生对更通用的智能至关重要的逻辑推理和常识理解。

As researchers wrestle with the challenges ahead, the ideas advocated by the symbolic AI camp are undergoing a kind of restoration. Nearly everyone acknowledges that the problems the symbolists tried, but largely failed, to solve must be addressed if artificial intelligence is to move forward. With the exception of a relatively small number of deep learning purists, many of whom seem to be associated with OpenAI, there is little confidence among researchers that simply scaling existing neural algorithms to take advantage of faster hardware and more data will be sufficient to produce the kind logical reasoning and common sense understanding that are essential for more general intelligence.

好消息是，这一次，我们可能看到的不是符号哲学和联结哲学之间的竞争，而是和解和整合的努力。这个新兴的研究领域被称为“神经符号人工智能”，可能代表人工智能未来最重要的举措之一。随着数十年有时激烈的竞争逐渐成为历史，新一代人工智能研究人员似乎愿意尝试弥合这两种方法之间的差距。麻省理工学院-IBM 沃森人工智能实验室主任大卫·考克斯 (David Cox) 表示，年轻的研究人员“根本没有那种历史”，“他们乐于探索交叉点。他们只想用人工智能做一些很酷的事情。” ⁴³

The good news is that this time, rather than competition between the symbolic and connectionist philosophies, we may see a reconciliation and an effort at integration. This emerging field of research has been dubbed “neuro-symbolic AI” and may represent one of the most important initiatives for the future of artificial intelligence. As decades of sometimes acrimonious competition fade into history, a new generation of AI researchers seems to be willing to try bridging the gap between the approaches. David Cox, the director of the MIT-IBM Watson AI Lab in Cambridge, MA, says that younger researchers “just don’t have any of that history” and “are happy to explore intersections. They just want to do something cool with AI.”⁴³

关于如何实现这种整合，有两种普遍的思想流派。最直接的方法可能是简单地构建混合系统，将神经网络与使用传统编程技术构建的软件模块相结合。能够处理逻辑和符号推理的算法将以某种方式与深度神经网络联系起来，重点是学习。这是 Elemental Cognition 的 David Ferrucci 团队正在实施的策略。第二种方法是找到一种方法，将符号 AI 功能直接实现到神经网络的架构中。这可能通过将必要的结构设计到深度神经网络中来实现，或者，我认为更具推测性的是，通过设计一个深度学习系统和一种非常有效的训练方法，使必要的结构以某种方式自然而然地出现。虽然年轻的研究人员可能愿意考虑所有可能性，但在那些职业生涯更成熟的研究人员中，关于最佳前进方向的激烈争论仍在继续。

There are two general schools of thought on how this integration might be accomplished. The most straightforward way may be to simply build hybrid systems that combine neural networks with software modules built using traditional programming techniques. Algorithms capable of handling logical and symbolic reasoning would somehow be linked to deep neural networks with a focus on learning. This is the strategy being pursued by David Ferrucci’s team at Elemental Cognition. The second approach would be to find a way to implement symbolic AI capabilities directly into the architecture of neural networks. This might be achieved by engineering the necessary structure into deep neural networks or, I think much more speculatively, by designing both a deep learning system and a training methodology so effective that the necessary structure would somehow emerge organically. While younger researchers may be willing to consider all possibilities, among those with more established careers there continues to be a sharp debate over the best way forward.

混合方法最直言不讳的倡导者之一是加里·马库斯 (Gary Marcus)，他直到最近还是纽约大学的心理学和神经科学教授。马库斯一直严厉批评他认为过分强调深度学习的做法，并撰写文章并参与辩论，他认为深度神经网络注定会保持浅薄和脆弱，除非将来自符号人工智能的想法直接注入其中，否则更通用的智能不太可能出现。马库斯在他的研究生涯中花了大量时间研究儿童如何学习和掌握语言，他认为纯深度学习方法几乎不可能接近人类儿童的非凡能力。他的批评并不总是受到深度学习社区的欢迎，尽管他与他人共同创办了一家机器学习初创公司，该公司于 2015 年被优步收购，但他被视为局外人，没有为该领域做出重大贡献。

One of the most outspoken advocates of a hybrid approach has been Gary Marcus, who until recently was a professor of psychology and neuroscience at New York University. Marcus has been a harsh critic of what he feels is an overemphasis on deep learning, and has penned essays and engaged in debates in which he argues that deep neural networks are destined to remain shallow and brittle, and that more general intelligence is very unlikely to emerge unless ideas drawn from symbolic AI are directly injected into the mix. Marcus spent much of his research career studying how children learn and acquire language and sees very little potential for a pure deep learning approach to come close to matching the remarkable capabilities of a human child. His criticisms have not always been well received by the deep learning community, where despite co-founding a machine learning startup company that was acquired by Uber in 2015, he is viewed as an outsider and someone who has not made significant contributions to the field.

一般来说，在深度学习方面投入最多的资深研究人员往往对混合方法不屑一顾。Yoshua Bengio 告诉我，目标应该是“解决传统人工智能试图解决的一些相同问题，但使用来自深度学习的构建模块”。44 ^Geoff Hinton 对这个想法更加不屑一顾，他说他“不相信混合动力车是答案”，并将这种系统比作鲁布·戈德堡式的混合动力汽车，其中电动机用于将汽油注入内燃机。45^问题是，到目前为止，还没有明确的策略将符号人工智能功能整合到完全由神经网络构建的系统中。正如 Marcus 指出的那样，深度学习的许多最杰出的成就，包括 Deep Mind 的 AlphaGo 系统，实际上都是混合系统，因为它们的成功仅仅依靠的是传统搜索算法和深度神经网络。

Generally, experienced researchers with the most invested in deep learning tend to be dismissive of the hybrid approach. Yoshua Bengio told me that the goal should be to “solve some of the same problems that classical AI was trying to solve but using the building blocks coming from deep learning.”⁴⁴ Geoff Hinton is even more disparaging of the idea, saying he doesn’t “believe hybrids are the answer” and comparing such a system to a Rube Goldberg–like hybrid car in which the electric motor is used to inject gasoline into an internal combustion engine.⁴⁵ The problem is that there is, so far, no clear strategy for incorporating symbolic AI capabilities into a system built entirely from neural networks. As Marcus points out, many of deep learning’s most prominent accomplishments, including Deep Mind’s AlphaGo system, are in fact hybrid systems because they succeeded only by relying on traditional search algorithms in addition to deep neural networks.

在研究人员争论混合模型的有效性的同时，另一个争论焦点是机器学习系统中固有结构的重要性。尽管深度神经网络确实经常包含某种程度的预先设计的结构——用于图像识别的卷积架构就是一个例子——但许多深度学习的铁杆支持者认为，这种结构应该保持在最低限度，而且这项技术将能够从接近空白的状态发展。例如，Yann LeCun 告诉我，“从长远来看，我们不需要精确的特定结构”，并指出人类大脑中没有这种神经结构的证据，他指出，“无论是视觉皮层还是前额叶皮层，皮层的微观结构似乎都非常非常统一。” ⁴⁶这一阵营的研究人员普遍认为，创新应该集中在开发改进的训练技术上，以提高相对通用的神经网络实现更大理解的能力。

As researchers argue over the efficacy of hybrid models, a parallel debate is focused on the importance of the innate structure built into machine learning systems. Though deep neural networks do often incorporate some degree of pre-designed structure—the convolutional architectures used for image recognition are one example—many hard-core deep learning advocates believe this should be kept to a minimum and that the technology will be capable of advancing from something fairly close to a blank slate. Yann LeCun, for example, told me that “in the long run we won’t need precise specific structures” and points out that there is no evidence of such neural structure in the human brain, noting that “the microstructure of the cortex seems to be very, very uniform all over, whether you’re looking at the visual or prefrontal cortex.”⁴⁶ Researchers in this camp generally argue that innovation should be focused on developing improved training techniques that boost the ability of relatively generic neural networks to achieve greater understanding.

像马库斯这样的研究人员，有着研究儿童认知发展的背景，强烈反对“白板”哲学。幼儿的大脑显然具有内置能力，有助于启动进一步的学习。出生几天后，婴儿就能识别人脸。在动物世界的其他方面，不依赖于学习的可操作智能的存在更加明显。冷泉港实验室的神经科学家安东尼·扎多尔指出，“松鼠出生几个月后就能从一棵树跳到另一棵树，小马驹几小时就能走路，蜘蛛生来就能捕猎。” ⁴⁷加里·马库斯经常提到阿尔卑斯山野山羊，这是一种生活在陡峭、危险地形上的野山羊。新生的野山羊能够在几个小时内站起来并在斜坡上行走，而任何通过反复试验的学习都意味着死亡。这是即插即用的技术：开箱即用。该阵营的研究人员认为，更通用、更灵活的人工智能同样需要内置认知机制，要么直接注入神经网络结构，要么通过混合方法集成。

Researchers like Marcus, with a background in the study of cognitive development in children, push back aggressively against the “blank slate” philosophy. The brains of young children clearly include built-in capability that helps to jump-start further learning. Within days of birth, babies are able to recognize human faces. Elsewhere in the animal world, the presence of actionable intelligence that does not rely on learning is even more obvious. Anthony Zador, a neuroscientist at Cold Spring Harbor Laboratory, points out that “a squirrel can jump from tree to tree within months of birth, a colt can walk within hours, and spiders are born ready to hunt.”⁴⁷ Gary Marcus often references the Alpine ibex, a species of mountain goat that lives out most of its life on steep, treacherous terrain. Newly born ibex are able to stand and navigate the slopes within hours in an environment where any kind of learning by trial and error would mean certain death. This is plug-and-play technology: it works out of the box. Researchers in this camp believe that more general, flexible artificial intelligence will likewise require built-in cognitive machinery, either injected directly into the structure of neural networks or integrated via a hybrid approach.

深度学习的倡导者有时会认为，尽管这种先天结构最终可能很重要，但它很可能作为持续学习过程的一部分自然而然地出现。然而，如果我们从生物大脑中寻找灵感，在我看来，大脑中的任何结构都不可能来自长期学习。我们知道，动物一生中的学习确实会在一定程度上重塑大脑；例如，人们常说“一起激发，一起连接”的神经元。问题是，单个生物体无法将其在一生中通过学习而形成的神经结构传递给后代。没有能力学习一些东西，然后以某种方式导致描述与该学习相关的大脑结构的信息被喷射到动物卵子或精子细胞的遗传密码中。个体生命中发展的任何大脑结构都会随着该生物体的消亡而消亡。因此，似乎很明显，大脑中的任何结构都必须是正常进化过程的结果，或者换句话说，随机突变，在极少数情况下，使生物体更能在环境中茁壮成长，因此更有可能传递给后代。一种方法可能是通过使用进化或遗传算法直接复制这个过程。然而，直接设计必要的结构可能是取得进展的更快方法。

Deep learning advocates sometimes suggest that, though such innate structure may ultimately be important, it is likely to arise organically as part of a sustained learning process. However, if we look to the biological brain for inspiration, it does seem to me that any structure in the brain cannot be the result of long-term learning. We know that learning over an animal’s lifetime does restructure the brain to some degree; it’s often said that neurons that “fire together, wire together,” for example. The problem is that there is no way for an individual organism to pass neural structure developed through learning during the course of its lifespan to its offspring. There is no ability to learn some stuff and then somehow cause information describing the brain structure associated with that learning to be ejected into the genetic code in the animal’s egg or sperm cells. Whatever brain structure develops within an individual life dies with that organism. Therefore, it seems clear that any structure in the brain must have resulted from the normal evolutionary process, or in other words random mutations that, in rare instances, make an organism more able to thrive in its environment and are therefore more likely to be passed along to offspring. One path might be to directly copy this process through the use of evolutionary or genetic algorithms. However, directly engineering the necessary structures may be a much faster way to achieve progress.

在混合方法与纯神经方法的争论中，你可能会说深度学习的拥护者有最终的反驳。人类大脑显然没有某种单独的计算机来运行特殊算法来完成其神经网络无法处理的所有事情。它只是神经元。不过，在我看来，混合方法可能会产生更多的近期实际结果。虽然纯神经实现显然是生物进化开辟的道路，但这不应该让我们对使用其他技术更快进步的可能性视而不见。也不应该仅仅因为它们被认为不够优雅而被拒绝可行的方法。当我们登陆月球时，我们并没有一艘科幻宇宙飞船，它只是快速下降、着陆然后再次起飞。相反，我们有一个更复杂——你甚至可以说笨重的——装置，包括一个登月舱和许多必须在途中丢弃的部件。也许有一天我们会拥有科幻宇宙飞船，但与此同时，我们已经登陆月球了。

In the debate over a hybrid versus a pure neural approach, you might say that deep learning adherents have the ultimate retort. The human brain clearly does not have some separate computer that runs special algorithms to do all the stuff that can’t be handled by its neural network. It’s just neurons all the way down. Still, to me it seems that the hybrid approach may be likely to produce more near-term practical results. While a purely neural implementation is clearly the path that was forged by biological evolution, this should not blind us to the possibility of faster progress using other techniques. Nor should viable approaches be dismissed simply because they are perceived as inelegant. When we landed on the moon, we didn’t have a science fiction spaceship that simply zoomed down, landed and then took off again. Rather, we had a much more complicated—you might even say clunky—contraption involving a lunar module and many parts that had to be discarded along the way. Someday perhaps we’ll have the science fiction spacecraft, but in the meantime we have landed on the moon.

通用机器智能发展道路上的一些关键挑战

SOME KEY CHALLENGES ON THE PATH TO GENERAL MACHINE INTELLIGENCE

大多数人工智能研究人员都认识到，要实现接近人类水平的人工智能，需要取得重大突破，但对于哪些挑战最为重要，或者哪些挑战应该首先攻克，并没有达成广泛的共识。Yann LeCun 经常使用穿越山脉的比喻。只有登上第一座山峰，你才能看到背后的障碍。需要克服的障碍与构建能够真正理解自然语言并进行有意义、不受约束的对话的机器的目标重叠且不可避免地相交。让我们更详细地看看人工智能研究需要解决的一些关键挑战。这个列表并非详尽无遗，但克服这些障碍的机器智能将比当今存在的任何智能更接近 AGI。同样，一个真正精通解决这些挑战的系统可能会产生具有巨大商业和科学价值的实际应用。

Most AI researchers recognize that significant breakthroughs will be required in order to achieve something close to human-level artificial intelligence, but there is no broad agreement on precisely what challenges are most important, or which ones should be attacked first. Yann LeCun often uses an analogy of navigating a mountain range. Only after you climb the first peak will you be able to see the obstacles that lie behind it. The hurdles that will need to be surmounted overlap and invariably intersect with the goal of building machines with the ability to truly understand natural language and engage in meaningful, unconstrained conversation. Let’s look in a bit more detail at some of the key challenges that AI research will need to address. This list is not intended to be exhaustive, but a machine intelligence that cleared these hurdles would be dramatically closer to AGI than anything that exists today. Likewise, a system that was truly proficient in addressing any one of these challenges would likely spawn practical applications with enormous commercial and scientific value.

常识推理

Common Sense Reasoning

我们所说的常识本质上是对世界及其运作方式的共同了解。我们生活的几乎每个方面都依赖常识，但它对我们的交流方式尤其重要。常识填补了未说出口的空白，让我们能够通过省略大量支持信息来极大地浓缩我们的语言。

What we refer to as common sense essentially amounts to a shared knowledge of the world and the way it works. We rely on common sense in nearly every aspect of our lives, but it is especially important to the way we communicate. Common sense fills in the unspoken gaps and allows us to dramatically condense our language by leaving out massive amounts of supporting information.

虽然几乎任何成年人都能毫不费力地利用这些内置知识，但事实证明，对机器来说，做到这一点是一个巨大的挑战。将常识融入人工智能是一个与符号人工智能与纯神经方法的争论密切相关的目标，也与将结构和知识融入人工智能系统的需求密切相关。

While virtually any adult is able to effortlessly draw on this built-in body of knowledge, doing the same has proven to be an enormous challenge for machines. Imbuing artificial intelligence with common sense is an objective that is deeply intertwined with the debates over symbolic AI versus a pure neural approach, as well as the need for structure and knowledge to be engineered into AI systems.

近年来，人工智能系统取得了重要进展，它们可以分析文本并正确回答有关材料的问题。例如，2018 年 1 月，微软和中国科技巨头阿里巴巴合作开发的软件在斯坦福大学研究人员设计的阅读理解测试中表现略优于人类平均水平。48^斯坦福测试提出的问题基于维基百科文章，其中正确答案由人工智能系统直接从“阅读”的文章中提取的一段文本组成。换句话说，我们看到的不是真正的理解，而是信息提取和模式识别的表现——正如我们所见，深度学习系统非常擅长这一点。当问题需要任何程度的常识推理或依赖对世界的隐性知识时，此类测试的表现就会急剧下降。

Recent years have seen important progress in AI systems that can analyze text and then correctly answer questions about the material. In January 2018, for example, software created by a collaboration between Microsoft and the Chinese tech giant Alibaba was able to slightly outperform the human average on a reading comprehension test created by researchers at Stanford University.⁴⁸ The Stanford test presents questions based on Wikipedia articles, in which the correct answer consists of a span of text drawn directly from the article “read” by the AI system. In other words, what we’re seeing is a demonstration not of true comprehension, but rather of information extraction and pattern recognition—something that, as we’ve seen, deep learning systems are extraordinarily good at. When the questions require any degree of common sense reasoning or reliance on implicit knowledge of the world, performance on such tests falls off dramatically.

要了解人工智能系统在常识方面遇到的困难，最好的方法之一就是研究特殊格式的句子，即 Winograd 模式。这些句子由斯坦福大学计算机科学教授 Terry Winograd 开发，利用模糊代词的力量来测试机器智能运用常识推理的能力。

One of the best ways to understand the struggle that AI systems have with common sense is to look at specially formatted sentences known as Winograd schemas. Developed by Terry Winograd, a computer science professor at Stanford, these sentences leverage the power of ambiguous pronouns to test a machine intelligence’s ability to employ common sense reasoning.

以下是示例：⁴⁹

Here’s an example:⁴⁹

由于担心发生暴力事件，市议会拒绝向示威者发放许可。

The city council refused the demonstrators a permit because they feared violence.

谁害怕暴力？答案对几乎所有人来说都很简单：市议会。

Who feared violence? The answer is easy for virtually anyone: the city council.

但现在只需改变句子中的一个词：

But now change just one word in the sentence:

由于示威者提倡暴力，市议会拒绝向他们发放许可。

The city council refused the demonstrators a permit because they advocated violence.

谁主张暴力？

Who advocated violence?

将“害怕”改为“提倡”会完全改变代词“他们”的含义。仅从句子中提取信息是无法正确回答这个问题的。你必须了解一些关于世界的事情，特别是市议会更喜欢和平的街道，而愤怒的示威者可能倾向于暴力。

Changing “feared” to “advocated” completely shifts the meaning of the pronoun “they.” There is no way to answer this question correctly simply by extracting information from the sentence. You have to understand something about the world, specifically that a city council would prefer peaceful streets, while angry demonstrators might be inclined to violence.

这里还有一些其他的例子，其中括号中显示了改变句子含义的替代词：

Here are some other examples with the alternate words that shift the meaning of the sentence shown in brackets:

这个奖杯太小/大了，放不进棕色的手提箱。

The trophy doesn’t fit into the brown suitcase because it’s too [small/large].

什么太[小/大]了？

What is too [small/large]?

由于速度太快/太慢，送货卡车从校车旁边飞驰而过。

The delivery truck zoomed by the school bus because it was going so [fast/slow].

什么事情发生得这么快/慢？

What was going so [fast/slow]?

走到楼梯口后，汤姆把书包扔给了雷。

Tom threw his schoolbag down to Ray after he reached the [top/bottom] of the stairs.

谁到达了楼梯的[顶部/底部]？

Who reached the [top/bottom] of the stairs?

对于这类问题，任何正常运作、识字的成年人都可能获得接近完美的分数。因此，及格分数的门槛应该设得很高。然而，面对一串 Winograd 模式，最好的计算机算法也只比随机猜测好一点点。

For a series of questions like these, any normally functioning, literate adult would likely achieve a score very close to perfection. Therefore, the threshold for a passing grade should be set very high. Faced with a list of Winograd schemas, however, the best computer algorithms perform only marginally better than random guessing.

华盛顿西雅图的艾伦人工智能研究所正在开展一项旨在将常识融入机器智能的最重要的计划。艾伦研究所首席执行官奥伦·埃齐奥尼告诉我，这项名为“马赛克计划”的计划部分源于该研究所对微软联合创始人保罗·艾伦的愿景的追求，即开发一个能够阅读科学教科书章节并回答章节末尾问题的人工智能系统。埃齐奥尼告诉我，尽管他的团队为实现这一目标所做的尝试是“最先进的”，但结果并不理想，通常成绩只有 D 左右。最大的障碍之一是回答问题时处理常识和逻辑推理的能力。例如，人工智能系统可以很容易地从生物学教科书中学习有关光合作用的事实材料。但埃齐奥尼说，真正的挑战在于当你有这样的问题时：“如果你在一个黑暗的房间里养了一株植物，然后你把它移到靠近窗户的地方，植物的叶子会长得更快、更慢还是一样快？” ⁵⁰这就需要你明白靠近窗户的地方光线更充足，并且能够推断出这将使植物生长得更快。

One of the most important initiatives geared toward building common sense into machine intelligence is occurring at the Allen Institute for AI in Seattle, Washington. Oren Etzioni, the CEO of the Allen Institute, told me that this effort, dubbed Project Mosaic, grew in part out of the Institute’s quest to fulfill Microsoft co-founder Paul Allen’s vision for an AI system that could read a chapter in a science textbook and then answer the questions at the end of the chapter. Etzioni told me that while his team’s attempts to accomplish this were “state of the art,” the results were less than stellar, typically resulting in a grade of around a D. One of the major stumbling blocks was the ability to handle common sense and logical reasoning while answering the questions. It’s fairly easy for an AI system to learn factual material about photosynthesis from a biology textbook, for example. But the real challenge, says Etzioni, is when you have a question like “If you have a plant in a dark room and you move it nearer the window, will the plant’s leaves grow faster, slower or at the same rate?”⁵⁰ This requires understanding that there will be more light closer to the window and the ability to reason that this will allow the plant to grow faster.

Mosaic 项目的第一个目标是创建一套标准基准，用于衡量机器展现常识的能力。一旦完成，该研究所计划部署各种技术，包括“众包、自然语言处理、机器学习和机器视觉” ^51，以生成内置的世界知识，这些知识将是让人工智能系统具备常识所必需的。

The first goal of Project Mosaic is to create a standard set of benchmarks designed to measure the ability of machines to exhibit common sense. Once this is complete, the Institute plans to deploy a variety of techniques, including “crowdsourcing, natural language processing, machine learning, and machine vision”⁵¹ to generate the built-in knowledge of the world that will be required to imbue an AI system with common sense.

尽管 Etzioni 和他的团队坚信使用一种将多种技术结合在一起的混合方法，但正如你所预料的那样，这个想法在最坚定的深度学习支持者中并没有引起多大热情。当我问 Yoshua Bengio 他是否认为 Project Mosaic 这样的努力很重要，或者他是否认为常识推理可能会以某种方式从学习过程中自然而然地出现时，他毫不怀疑他对深度学习方法的信心：“我相信常识会作为学习过程的一部分出现。它不会因为有人把一点知识塞进你的脑袋里而出现，这不是人类的工作方式。” ⁵² Yann LeCun 同样认为获得常识的途径是通过学习，他告诉我 Facebook 的人工智能研究团队正在研究“让机器通过观察不同的数据源来学习——学习世界是如何运作的。我们正在建立一个世界模型，也许某种形式的常识会出现，也许这个模型可以用作一种预测模型，让机器像人一样学习。” ⁵³

While Etzioni and his team are strong believers in using a hybrid approach that will bring together a variety of techniques, this idea, as you might expect, generates little enthusiasm among the staunchest deep learning advocates. When I asked Yoshua Bengio if he thought efforts like Project Mosaic were important, or if he thought common sense reasoning might somehow emerge organically from the learning process, he left no doubt about his faith in a deep learning approach: “I’m sure common sense will emerge as part of the learning process. It won’t come up because somebody sticks little bits of knowledge into your head, that’s not how it works for humans.”⁵² Yann LeCun likewise believes the path to common sense is through learning, telling me that Facebook’s AI research team is working on “getting machines to learn by observation from different data sources—learning how the world works. We’re building a model of the world so that perhaps some form of common sense will emerge and perhaps that model could be used as kind of a predictive model that would allow a machine to learn the way people do.”⁵³

好消息是，一些世界上最聪明的人工智能研究人员正在积极研究这两种方法。如果人工智能系统能够可靠地部署我们认为人类理所当然的常识推理，那么这将是一项非凡的进步，无论它是自然产生的还是通过更工程化的方法产生的。

The good news is that both approaches are being pursued aggressively by some of the world’s brightest AI researchers. A breakthrough that results in an AI system that is able to reliably deploy the kind of common sense reasoning that we take for granted in humans would be an extraordinary advance, regardless of whether it emerges organically or results from a more engineered approach.

无监督学习

Unsupervised Learning

正如我们所见，用于训练深度学习系统的两种主要技术是监督学习和强化学习，前者需要大量标记数据，后者需要大量反复试验才能使算法成功完成任务。虽然人类同样使用这些技术，但它们只占幼儿大脑学习的一小部分。幼儿通过简单的观察、倾听父母的声音以及直接与周围世界互动和实验来学习。

As we’ve seen, the two primary techniques used to train deep learning systems are supervised learning, which requires large quantities of labeled data, and reinforcement learning, which requires a huge number of trial-and-error iterations as an algorithm attempts to succeed at a task. Though human beings likewise employ these techniques, they constitute only a tiny fraction of the learning that goes on in a young child’s mind. Very young children learn from simple observation, by listening to the voices of their parents and by engaging and experimenting directly with the world around them.

新生儿几乎立刻就开始了这一过程，早在他们具备以任何有意识的方式与周围环境互动的身体能力之前，他们就直接从周围环境中学习。不知何故，他们设法开发出一个物理世界模型，并开始构建常识所依赖的知识基础。这种无需借助结构化和标记数据直接学习的能力被称为“无监督学习”。这种非凡的能力很可能是由儿童大脑中内置的某种认知结构实现的，但毫无疑问，人类儿童独立学习的能力，尤其是学习语言的能力，远远超过最强大的深度学习系统所能实现的任何能力。

Newly born babies begin the process almost immediately, learning directly from their environment long before they have the physical ability to interact with it in any deliberate way. Somehow, they manage to develop a physical model of the world and begin to build the base of knowledge that underlies common sense. This ability to learn directly and without assistance from structured and labeled data is known as “unsupervised learning.” This remarkable ability may well be enabled by some kind of cognitive structure built into the child’s brain, but there is no doubt that the ability of a human child to learn independently, and especially to acquire language, vastly outpaces anything that can be accomplished with the most powerful deep learning system.

这种早期的无监督学习随后支持了以后更高级的知识获取。即使年龄较大的孩子的学习在某种程度上受到监督，所需的训练数据也只是最先进算法所需数据的一小部分。深度神经网络可能需要数千张带标签的训练照片，才能可靠地将动物名称附加到图像上。相比之下，父母指着一只动物说一次“这是一只狗”就足够了。一旦孩子能够识别出这种动物，她很可能在任何情况下都能做到这一点；狗可能是坐着、站着或在马路上跑着，孩子仍然可以始终如一地给它起个名字。

This early unsupervised learning then supports more advanced knowledge acquisition later on. Even when an older child’s learning is to some extent supervised, the training data required is a tiny fraction of what would need to be provided to even the most advanced algorithm. A deep neural network might require many thousands of labeled training photographs before it can reliably attach the names of animals to their images. In contrast, a parent pointing to an animal and saying, “This is a dog” a single time might well be sufficient. And once the child can identify the animal, she can likely do so in any configuration; the dog could be sitting or standing or running across the road, and still the child can consistently attach a name to it.

无监督学习是目前人工智能领域最热门的研究课题之一。谷歌、Facebook 和 DeepMind 都有团队专注于这一领域。然而，进展缓慢，迄今为止几乎没有出现真正实用的应用。事实是，没有人真正知道人类大脑如何实现其无与伦比的从非结构化数据中自主学习的能力。目前大多数研究都集中在听起来不那么雄心勃勃的无监督学习变体上，例如预测学习或自我监督学习。示例项目可能包括尝试预测句子中的下一个单词或视频中下一帧的组成图像。虽然这些任务似乎与人类管理的相去甚远，但许多研究人员认为，预测能力是智能的绝对核心，而像这样的实验将推动事情朝着正确的方向发展。无监督机器学习的真正突破的重要性怎么说都不为过。例如，Yann LeCun 认为，这很可能是通向通用智能几乎所有其他方面进步的大门，他说：“在我们弄清楚如何做到这一点之前……我们不会取得重大进展，因为我认为这是学习足够多的世界背景知识的关键，这样常识就会出现。这是主要的障碍。” ⁵⁴

Unsupervised learning is currently one of the hottest research topics in the field of artificial intelligence. Google, Facebook and DeepMind all have teams focused in this area. Progress, however, has been slow, and few if any truly practical applications have so far emerged. The truth is that no one really has any idea exactly how the human brain achieves its unparalleled competence at autonomously learning from unstructured data. Most current research is focused on less ambitious-sounding variants of unsupervised learning, such as predictive learning or self-supervised learning. Example projects might include trying to predict the next word in a sentence or the image that makes up the next frame in a video. While these kinds of tasks may seem a far cry from what humans manage, many researchers believe that the ability to make predictions is absolutely central to intelligence and that experiments like these will drive things in the right direction. It’s difficult to overstate the magnitude of a genuine breakthrough in unsupervised machine learning. Yann LeCun, for example, believes it may well be the gateway that leads to progress on nearly every other aspect of general intelligence, saying, “Until we figure out how to do this… we’re not going to make significant progress because I think that’s the key to learning enough background knowledge about the world so that common sense will emerge. That’s the main hurdle.”⁵⁴

理解因果关系

Understanding Causation

学习统计学的学生经常被提醒“相关性不等于因果关系”。对于人工智能，尤其是深度学习系统，理解止于相关性。加州大学洛杉矶分校著名计算机科学家 Judea Pearl 在过去三十年中彻底改变了因果关系的研究，并构建了一种表达因果关系的正式科学语言。2011 年获得图灵奖的 Pearl 喜欢指出，虽然任何人都可以直观地理解日出导致公鸡打鸣，而不是反之亦然，但最强大的深度神经网络可能无法获得类似的洞察力。因果关系不能仅通过分析数据来得出^。55

Students studying statistics are often reminded that “correlation does not equal causation.” For artificial intelligence, and especially deep learning systems, understanding ends at correlation. Judea Pearl, a renowned computer scientist at UCLA, has over the past thirty years revolutionized the study of causation and constructed a formal scientific language for expressing causal relationships. Pearl, who won the Turing Award in 2011, likes to point out that while any human understands intuitively that the sunrise causes a rooster to crow, rather than vice versa, the most powerful deep neural network would likely fail to achieve a similar insight. Causation cannot be derived simply by analyzing data.⁵⁵

人类拥有独特的能力，不仅可以检测相关性，还可以理解因果关系，而且我们只需极少的例子就能做到这一点。麻省理工学院计算认知科学教授 Joshua Tenenbaum 将自己的研究重点描述为“逆向工程人类思维”，希望获得有助于构建更智能的人工智能系统的见解，他指出：

Human beings have a unique ability to not just detect correlations but also to understand causal effects, and we can do so on the basis of remarkably few examples. Joshua Tenenbaum, a professor of computational cognitive science at MIT, who describes his research focus as “reverse engineering the human mind” in the hope of gaining insights that will be useful in building smarter AI systems, points out:

即使是年幼的孩子也常常能从一个或几个例子中推断出新的因果关系——他们甚至不需要看足够多的数据就能发现统计上显著的相关性。想想你第一次看到智能手机的时候，无论是 iPhone 还是其他带有触摸屏的设备，当有人用手指在一块小玻璃板上滑动时，突然有东西亮了起来或移动了。你以前从未见过这样的事情，但你只需要看一两次就能明白存在这种新的因果关系，然后这只是你学习如何控制它并完成各种有用事情的第一步^。56

Even young children can often infer a new causal relation from just one or a few examples—they don’t even need to see enough data to detect a statistically significant correlation. Think about the first time you saw a smartphone, whether it was an iPhone or some other device with a touchscreen where somebody swipes their finger across a little glass panel, and suddenly something lights up or moves. You had never seen anything like that before, but you only need to see that once or a couple of times to understand that there’s this new causal relation, and then that’s just your first step into learning how to control it and to get all sorts of useful things done.⁵⁶

理解因果关系对于想象力和心理反事实情景的生成至关重要，而这些情景使我们能够解决问题。与需要失败数千次才能找出成功方法的强化学习算法不同，我们可以在头脑中运行一种模拟，并探索替代行动方案的可能结果。如果没有对因果关系的直观理解，这是不可能的。

Understanding causal relationships is critical to imagination and to the generation of mental counterfactual scenarios that enable us to solve problems. Unlike a reinforcement learning algorithm that needs to fail thousands of times before figuring out how to succeed, we can run a kind of simulation in our heads and explore the likely outcomes of alternate courses of action. This would be impossible without an intuitive grasp of causation.

Pearl 和 Tenenbaum 等研究人员认为，对因果关系的理解（本质上是提出和回答“为什么”的能力）将成为构建更通用的机器智能的重要因素。Pearl 在因果关系方面的研究对自然科学和社会科学产生了巨大影响，但他认为，人工智能研究人员大多未能领会这一观点，而且通常过于关注机器学习系统能够有效识别的相关性。57^然而，这种情况正在改变。例如，蒙特利尔大学的 Yoshua Bengio 和他的团队最近发表了一项重要研究，探讨了一种在深度学习系统中建立因果关系理解的创新方法^。58

Researchers like Pearl and Tenenbaum believe that an understanding of causal relationships—in essence the ability to ask and answer the question “Why?”—will be an essential ingredient in building more general machine intelligence. Pearl’s work on causation has had an enormous impact in the natural and social sciences, but he believes that AI researchers have largely failed to get the memo and have generally been too focused on the correlations that are so efficiently identified by machine learning systems.⁵⁷ This is changing, however. For example, Yoshua Bengio and his team at the University of Montreal recently published important research on an innovative way to build an understanding of causation into deep learning systems.⁵⁸

迁移学习

Transfer Learning

哈佛大学政治学家兼教授格雷厄姆·艾利森因提出“修昔底德陷阱”这一短语而闻名。该术语指的是希腊历史学家修昔底德的《伯罗奔尼撒战争史》，该书记录了公元前五世纪斯巴达与新兴雅典之间的冲突。格雷厄姆认为，斯巴达和雅典之间的战争代表了一种仍然适用的历史原则。在他 2017 年出版的《注定一战》一书中，他认为美国和中国陷入了当代的修昔底德陷阱，随着中国实力和影响力的不断崛起，冲突很可能不可避免^。59

Graham Allison, a political scientist and professor at Harvard, is known for coining the phrase “Thucydides’s Trap.” The term references the Greek historian Thucydides’s History of the Peloponnesian War, which chronicles the conflict between Sparta and a newly ascendant Athens in the fifth century BC. Graham believes that the war between Sparta and Athens represents a kind of historical principle that remains applicable today. In his 2017 book Destined for War, he argues that the United States and China are caught in a contemporary Thucydides’s Trap and that as China continues to rise in power and influence, conflict may well be inevitable.⁵⁹

人工智能系统能否阅读《伯罗奔尼撒战争史》等历史文献，然后成功地将其学到的知识应用于当代地缘政治局势？做到这一点将达到人工智能道路上最重要的里程碑之一：迁移学习。学习一个领域的信息，然后成功地将其运用到其他领域的能力是人类智能的标志之一，对创造力和创新至关重要。如果更通用的机器智能要真正发挥作用，它必须做的不仅仅是回答本章末尾的问题；它需要能够将其学到的知识和它开发的任何见解应用于全新的挑战。在人工智能系统有希望实现这一点之前，它需要远远超越目前在深度神经网络中凝聚的肤浅理解水平，实现真正的理解。事实上，在各种领域和新情况下应用知识的能力可能是机器智能真正理解的唯一最佳测试。

Could an artificial intelligence system read a historical document like the History of the Peloponnesian War and then successfully apply what it learns to a contemporary geopolitical situation? To do so would be to reach one of the most important milestones on the path toward artificial general intelligence: transfer learning. The ability to learn information in one domain and then successfully leverage it in other domains is one of the hallmarks of human intelligence and is essential to creativity and innovation. If more general machine intelligence is to be genuinely useful, it will have to do more than simply answer the questions at the end of the chapter; it will need to be able to apply what it learns, and any insights it develops, to entirely new challenges. Before an AI system has any hope of accomplishing this, it will need to move well beyond the superficial level of understanding that currently coalesces within deep neural networks and achieve genuine comprehension. Indeed, the ability to apply knowledge in a variety of domains and in novel situations may turn out to be the single best test for true understanding in a machine intelligence.

迈向人类水平的人工智能之路

THE PATH TO HUMAN-LEVEL ARTIFICIAL INTELLIGENCE

我采访过的几乎所有人工智能研究人员都认为，人类级别的人工智能是可以实现的，而且终有一天会成为必然。在我看来，这似乎是合理的。毕竟，人类大脑从根本上来说是一种生物机器。没有理由相信生物智能有什么神奇之处，也没有理由相信某种大致相当的东西有朝一日无法在完全不同的媒介中实现。

Nearly all the AI researchers I have spoken to believe that human-level artificial intelligence is achievable and someday will be inevitable. To me, this seems reasonable. The human brain, after all, is fundamentally a biological machine. There is no reason to believe that there is anything magical about biological intelligence or that something broadly comparable couldn’t someday be instantiated in an entirely different medium.

事实上，硅基基板似乎比驱动人类大脑的生物湿件具有许多优势。电子信号在计算机芯片中的传播速度远高于在大脑中的传播速度，任何有朝一日能够与我们的推理和交流能力相媲美的机器都将继续享有计算机目前相对于我们的所有优势。机器智能将受益于完美的记忆，即使是过去发生的事件，也能够以惊人的速度计算和筛选和搜索大量数据。它还能够直接连接到互联网或其他网络，并利用几乎无限的资源；它可以毫不费力地与其他机器交谈，就像它掌握了与我们交谈一样。换句话说，人类级别的人工智能从一开始就在很多方面都比我们优越。

Indeed, a silicon-based substrate would seem to have many advantages over the biological wetware that powers the human brain. Electronic signals propagate at vastly higher speeds in computer chips than in the brain, and any machine that someday equaled our ability to reason and communicate would continue to enjoy all the advantages that computers currently have over us. A machine intelligence would benefit from flawless memory, even of events that occurred deep in the past, and would have the ability to calculate and to sift and search through enormous troves of data at fantastic speed. It would also be able to directly connect to the internet or to other networks and tap into virtually limitless resources; it would effortlessly talk to other machines, even as it mastered conversation with us. In other words, human-level AI, from it very inception, would in a great many ways be superior to us.

尽管几乎所有人都相信有朝一日能够到达这个目的地，但通往目的地的路线和到达时间仍然笼罩在深深的不确定性之中。到目前为止，进展基本上是渐进的。例如，2017 年底，DeepMind 发布了 AlphaZero，这是其围棋 AlphaGo 系统的更新版。AlphaZero 不再需要对数千场人类围棋比赛的数据进行监督学习，而是从一张白纸开始，仅凭与自己对弈的模拟游戏就能学会超人水平的下棋。该系统还能够接受其他挑战的训练，包括国际象棋和日本将棋。AlphaZero 很快击败了最优秀的专用国际象棋算法，证明了它是地球上最顶尖的国际象棋选手——当然，这些算法已经能够轻松击败最有能力的人类选手。戴米斯·哈萨比斯 (Demis Hassabis) 告诉我，AlphaZero 可能代表了“信息完全”游戏的通用解决方案，换句话说，在这种类型的挑战中，你获胜所需的所有信息都可以作为棋盘上的游戏棋子或屏幕上的像素来轻松获得。

Though faith in someday reaching this destination is nearly universal, the route that will take us there, and the time of arrival, remain shrouded in deep uncertainty. So far, progress has largely been incremental. For example, in late 2017, DeepMind released AlphaZero, an update to its Go-playing AlphaGo system. AlphaZero dispensed with the need for a supervised learning regimen on data from thousands of Go matches played by humans and instead began with essentially a blank slate, learning to play at superhuman levels purely on the basis of simulated games played against itself. The system also had the ability to be trained for other challenges, including chess and the Japanese game shogi. AlphaZero quickly demonstrated that it is the top chess-playing entity on the planet by defeating the very best dedicated chess-playing algorithms—which, of course, were already able to easily dispense with the most capable human players. Demis Hassabis told me that AlphaZero probably represents a general solution to “information complete” games, or in other words the type of challenges in which all the information you need to prevail is readily available as game pieces on a board or pixels on a screen.

当然，我们生活的现实世界远非信息完备。几乎所有我们希望在未来利用先进人工智能的最重要的领域都需要具备在不确定情况下运作的能力，以及处理大量信息隐藏或根本无法获取的情况。2019 年 1 月，DeepMind 再次展示了其进步，发布了 AlphaStar，这是一个专为玩战略视频游戏《星际争霸》而设计的系统。《星际争霸》模拟了三个不同的外星物种之间为争夺资源而展开的银河斗争，每个物种都由在线玩家实时控制。《星际争霸》不是一款信息完备的游戏；相反，玩家需要“侦察”才能发现对手活动的隐藏信息。这款游戏还需要在广阔的游戏空间内进行长期规划和管理资源。2018年 12 月，AlphaStar 在一场比赛中以 5-0击败了顶级职业《星际争霸》玩家，这是 DeepMind 团队的另一个首创^。60

The real world in which we live is, of course, far from information complete. Nearly all the most important areas in which we would someday like to leverage advanced artificial intelligence require the ability to operate under uncertainty and to deal with situations where vast amounts of information are hidden or simply unattainable. In January 2019, DeepMind again demonstrated progress with its release of AlphaStar, a system designed to play the strategy video game StarCraft. StarCraft simulates a galactic struggle for resources between three different extraterrestrial species, each of which is controlled in real time by an online player. StarCraft is not an information complete game; rather, players need to “scout” in order to discover hidden information about their opponents’ activities. The game also requires long-term planning and management of resources across a vast game space. In another first for DeepMind’s team, AlphaStar defeated a top professional StarCraft player 5-0 in a match conducted in December 2018.⁶⁰

虽然这些成就令人印象深刻，但它们仍远未克服将当今的人工智能系统限制在高度特定、狭窄领域的重大限制。例如，AlphaStar 必须使用监督和强化学习技术进行大量训练，才能扮演特定的外星物种。切换到具有不同相对优势的不同物种需要从头开始重新训练。同样，AlphaZero 可以轻松实现称霸世界的国际象棋或将棋能力，但如果不重新训练该游戏，该系统将无法在跳棋游戏中击败儿童。即使是人工智能研究最前沿的最强大系统也仍然很肤浅和脆弱。正如艾伦研究所的奥伦·埃齐奥尼 (Oren Etzioni) 喜欢指出的那样，如果这些系统中的任何一个知道房间着火了，它们都会继续不受干扰地继续游戏。61^没有常识，没有真正的理解。

Though these achievements are impressive, they still do not come close to overcoming the major limitations that confine today’s AI systems to highly specific, narrow domains. AlphaStar, for example, must be extensively trained, using supervised and reinforcement learning techniques, to play in the role of a particular alien species. Switching to a different species, with different relative strengths, requires re-training from scratch. Likewise, AlphaZero can easily achieve world-dominating chess or shogi capability, but the system would not be able to beat a child at a game of checkers without retraining on that game. Even the most powerful systems at the very forefront of AI research remain shallow and brittle. And as the Allen Institute’s Oren Etzioni likes to point out, any of these systems would continue to play unperturbed if they learned the room was on fire.⁶¹ There is no common sense, no true understanding.

克服这些限制并成功建造一台真正的思考机器需要多长时间？在我参与《智能建筑师》一书中记录的对话时，我对人工智能领域的顶尖人才进行了非正式调查。我要求与我交谈的 23 个人中的每一位预测通用人工智能实现的概率至少为 50%。大多数参与者要求匿名。与我交谈的五位研究人员拒绝做出任何预测，指出通往人类水平的人工智能的道路高度不确定，需要克服未知数量的具体挑战。尽管如此，世界上 18 位顶尖的人工智能专家还是给了我他们最好的猜测，我认为下表中显示的结果非常有趣^。62

How long might it take to overcome these limitations and succeed in building a genuine thinking machine? As I engaged in the conversations that are recorded in my book Architects of Intelligence, I conducted an informal survey of the top minds in the field of AI. I asked each of the twenty-three individuals I spoke with to give me his or her prediction for the year artificial general intelligence would have at least a fifty percent probability of being achieved. Most of the participants requested that their guesses remain anonymous. Five of the researchers I spoke with declined to make a prediction at all, pointing out that the path to human-level AI is highly uncertain and that an unknown number of specific challenges will need to be overcome. Nonetheless, eighteen of the world’s foremost AI experts did give me their best guess, and I think the results, which are shown in the table below, are very interesting.⁶²

AGI 实现年份

Year AGI Achieved

从 2021 年开始

Years from 2021

猜测次数

Number of Guesses

2029

1（库兹韦尔）

1 (Kurzweil)

2036

2038

2040

2068

四十七

2080

2088

2098

2118

2168

147

2188

167

2200

179

1（布鲁克斯）

1 (Brooks)

请注意，这些猜测是在 2018 年做出的，这解释了以“8”结尾的年份占多数的原因。例如，对 2038 年的猜测实际上是对“从现在起 20 年后”的猜测。我强烈怀疑，如果我今天再次让同一群人猜测，我会得到基本相同的估计值，将这些数字推迟三年左右。这确实引发了人们的担忧，即 AGI 的成就可能会成为物理学家经常讲的关于核聚变的老笑话的牺牲品：“它总是未来三十年。”

Note that these guesses were made in 2018, which explains the preponderance of years ending in “8.” A guess of 2038, for example, was actually a guess of “twenty years from now.” I strongly suspect that if I asked the same people to guess again today, I would get essentially the same estimates, moving these numbers out by three years or so. This does raise the concern that the achievement of AGI might fall prey to the old joke that physicists have often told about nuclear fusion: “It’s always thirty years in the future.”

平均猜测是 2099 年，或者大约 80 年后。i ^,⁶³ ^这些预测被两位愿意公开发表意见的人的猜测巧妙地括起来。正如我们所见，雷·库兹韦尔坚持认为人类水平的人工智能将在 2029 年出现——现在只剩 8 年的时间了。iRobot 公司的联合创始人、被广泛认为是世界顶尖机器人专家之一的罗德尼·布鲁克斯 (Rodney Brooks) 认为 AGI 需要近 180 年才能到来。我认为，预测之间的巨大差距——多位研究人员预计人类水平的人工智能将在十年或二十年内出现，而其他人则认为可能需要几个世纪——生动地说明了人工智能的未来是多么难以预测。

The average guess was the year 2099, or roughly eighty years from now.ⁱ^,⁶³ The predictions are neatly bracketed by guesses from two people who were willing to go on the record. Ray Kurzweil, as we’ve seen, remains adamant that human-level AI will come into existence by 2029—now just eight years away. Rodney Brooks, a co-founder of iRobot Corporation and widely regarded as one of the world’s foremost roboticists, thinks it will take nearly 180 years for AGI to arrive. This gaping chasm between predictions—with multiple researchers anticipating human-level AI within a decade or two, whereas others think it could be centuries—is, I think, a vivid illustration of just how unpredictable the future of artificial intelligence is likely to be.

我认为，构建人类级别的人工智能是人工智能领域最令人着迷的话题。有朝一日，它很可能会成为人类最具影响力和颠覆性的创新。然而，与此同时，人工智能作为一种实用工具仍将相对狭窄，而且在许多方面相当有限。可以肯定的是，随着该领域最前沿的研究被纳入其中，旨在解决现实问题的人工智能系统将不断升级，但在可预见的未来，这项新技术的力量将不是由单一的、高度灵活的机器智能来实现的，而是由大量特定应用来实现的，这些应用已经开始在工业、经济、社会甚至文化的几乎每个方面扩展。

THE QUEST TO build human-level AI is, I think, the single most fascinating topic in the field of artificial intelligence. Someday it may well result in humanity’s most consequential and disruptive innovation. In the meantime, however, artificial intelligence as a practical tool will remain relatively narrow, and in many ways quite limited. To be sure, AI systems designed to solve real-world problems will be continuously upgraded as research at the very frontier of the field is incorporated, but for the foreseeable future, the power of this new technology will be delivered not by a single, highly flexible machine intelligence, but rather by an explosion of specific applications that are already beginning to scale across nearly every aspect of industry, the economy, society and even culture.

毫无疑问，人工智能有可能带来巨大的利益，特别是在医疗保健、科学研究和广泛的技术创新等关键领域。然而，这项技术还有另一面。人工智能将带来前所未有的挑战和危险——对就业和经济、个人隐私和安全，甚至可能最终对我们的民主制度甚至文明本身造成威胁。这些风险将是接下来三章的主要关注点。

Without question, AI has the potential to deliver profound benefits, especially in critical areas like healthcare, scientific research and broad-based technical innovation. There is another side to the technology, however. Artificial intelligence will come coupled with unprecedented challenges and dangers—to jobs and the economy, to personal privacy and security and perhaps ultimately to our democratic system and even to civilization itself. These risks will be the primary focus of the next three chapters.

脚注

Footnote

i与已进行的其他调查相比，这一平均值较为悲观。这些调查包括大量经验水平各异的人工智能研究人员，调查对象通常是在人工智能会议上。大多数结果都显示，AGI 的出现时间集中在 2040 年至 2050 年左右，概率为 50%。请参阅第 5 章尾注 63，了解此类调查的列表。

i This average is pessimistic relative to other surveys that have been conducted. These have included a much larger number of AI researchers with widely different experience levels, often at AI conferences. Most results have clusters around the years 2040 to 2050 for AGI with a fifty percent probability. See endnote 63, Chapter 5, for a list of such surveys.

⁶³

第六章

CHAPTER 6 工作岗位消失和人工智能带来的经济后果

DISAPPEARING JOBS AND THE ECONOMIC CONSEQUENCES OF AI

我在 2015 年出版的《机器人崛起：技术与失业未来的威胁》一书中指出，人工智能和机器人技术的进步最终将摧毁大量常规和可预测的工作，这可能会导致不平等加剧和结构性失业。当我在 2020 年 1 月开始写这本书时，我认为本章中我面临的主要任务是在二战以来最长的经济复苏和约 3.6% 的总体失业率面前捍卫这一论点。

IN MY 2015 BOOK Rise of the Robots: Technology and the Threat of a Jobless Future, I argued that advances in artificial intelligence and robotics would eventually destroy a great many jobs that tended to be routine and predictable—potentially leading to increased inequality and structural unemployment. As I began writing this book in January 2020, I assumed that the main task before me in this chapter would be to defend that thesis in the face of the longest economic recovery since World War II and a headline unemployment rate of about 3.6 percent.

毋庸置疑，新冠病毒大流行以及随之而来的美国和全球经济停摆，将我们带入了一个全新的经济现实。尽管如此，我相信我在危机爆发前计划提出的论点仍然具有高度现实意义。即使在失业率处于历史低位的时期，我相信我在 《机器人崛起》中讨论的趋势仍然牢牢地发挥作用，而当前危机前几年经济指标所显示的相对繁荣至少在某种程度上是一种假象。在疫情之后，工作自动化程度提高的趋势可能会进一步加剧，并可能在我们期待从当前经济灾难中复苏时产生巨大影响。

Needless to say, the coronavirus pandemic and the ensuing shut down of economies in the United States and across the globe have led us into an entirely new economic reality. Nonetheless, I believe the arguments I planned to make before the crisis emerged remain highly relevant. Even in a time of historically low unemployment, I believe that the trends I discussed in Rise of the Robots remained firmly in play, and that the relative prosperity suggested by economic indicators in the years leading up to the current crisis was, at least to some extent, an illusion. In the wake of the pandemic, the trend toward increased job automation may well be amplified and could have a dramatic impact as we look forward to recovery from the current economic disaster.

假设你是 1965 年的美国经济学家。当你审视美国经济和就业市场时，你会发现大约 97% 的 25 至 54 岁男性（这个年龄足以完成学业但还不到退休年龄）要么在工作，要么在积极寻找工作。这对你来说似乎是完全可以预料到的，也很正常。现在假设一位来自未来的时间旅行者出现并告诉你，到 2019 年，只有大约 89% 的黄金工作年龄男性会进入劳动力市场，而到 2050 年，这个年龄段的美国男性中完全被剥夺就业权利的比例可能会增长到四分之一甚至三分之一^。i^,¹

Imagine that you are an American economist in the year 1965. As you gazed out over the U.S. economy and job market, you would see that about ninety-seven percent of men between the ages of twenty-five and fifty-four—old enough to have completed schooling but too young to retire—are either employed or actively seeking work. This would seem entirely expected and normal to you. Now suppose that a time traveler from the future appears and tells you that in the year 2019, only about eighty-nine percent of prime working age men will be in the workforce—and that by 2050 the fraction of American men in this age group who are completely disenfranchised from the job market might well grow to a quarter or even a third.ⁱ^,¹

似乎你肯定会对此感到担忧。也许你甚至会想到“大规模失业”这个词。你肯定会想知道那些失业的人到底在干什么。但现在，时间旅行者告诉你，政府报告的 2019 年总体失业率远低于 4%，利率低于 1965 年的水平。时间旅行者指出，这两项指标都接近历史低点。此外，你被告知，美联储并没有计划加息，而是暗示可能会进一步降低利率以提振经济。

It seems like a good bet that you would find this alarming. Perhaps the phrase “mass unemployment” would even cross your mind. You would surely wonder just what all those non-working men were up to. But now the time traveler tells you that the headline unemployment rate reported by the government in 2019 is significantly below four percent and that interest rates are below the 1965 level. Both measures, points out the time traveler, are close to historic lows. Furthermore, you are told, the U.S. Federal Reserve, rather than planning to raise interest rates, is signaling that it might well lower them further in an effort to boost the economy.

所有这些事情对于 20 世纪中后期的经济学家来说可能都是相当令人惊讶和困惑的。正如我们将在本章中看到的那样，美国以及许多其他发达国家的经济和就业市场现在的运作方式似乎违背了许多曾经似乎得到实证证据坚实支持的规则和假设。

All of these things would likely be quite surprising and confusing to an economist from the mid- to late twentieth century. As we’ll see in this chapter, the economy and job market in the United States, as well as in many other developed countries, are now operating in ways that seem to defy many of the rules and assumptions that once appeared to be solidly backed by empirical evidence.

在《机器人的崛起》一书中，我指出这些变化很大程度上是由信息技术的加速进步所推动的。如今，我们已经历了一系列关键创新——先进的工厂自动化、个人电脑革命、互联网、云计算和移动技术的兴起——而由此产生的变革已持续了数十年。然而，最重要的技术影响仍然在未来。人工智能的崛起有可能以比我们以前所见的任何影响都更为剧烈和根本的方式颠覆就业市场和我们的整体经济体系。

In Rise of the Robots, I argued that these changes are being driven largely by accelerating progress in information technology. A long list of key innovations is now behind us—advancing factory automation, the personal computer revolution, the internet, the rise of cloud computing and mobile technology—and the resulting transformation has been playing out over the course of decades. The most important technological impact, however, still lies in the future. The rise of artificial intelligence has the potential to upend both the job market and our overall economic system in ways that are far more dramatic—and foundational—than anything we have seen previously.

当我们站在即将到来的颠覆的最前沿时，我们有充分的理由感到担忧。过去一二十年发生的变革无疑在难以想象的政治动荡中发挥了重要作用，并撕裂了社会结构。例如，研究表明，美国最容易受到工作自动化影响的地区与 2016 年总统大选中强烈支持唐纳德·特朗普的选民之间存在直接关联。2^在冠状病毒大流行颠覆我们的生活之前，人们更多地关注另一场正在摧毁美国的健康危机，中产阶级大量失业的地区也往往处于阿片类药物泛滥的前线。3^如果我们迄今为止看到的变化与可能发生的事情相比相形见绌，那么未来就有可能出现前所未有的社会和经济混乱——以及更危险的政治煽动家的崛起，他们将利用这种迅速变化的形势必然伴随的恐惧而茁壮成长。

As we stand at the leading edge of the coming disruption, there are good reasons to be concerned. The transformations that have occurred over just the past decade or two have arguably played an important role in unimaginable political upheaval and have rended the very fabric of society. Studies, for example, have shown a direct correlation between regions in the United States most vulnerable to job automation and voters who strongly supported Donald Trump in the 2016 presidential election.² Before the coronavirus pandemic upended our lives, there was more focus on another health crisis that has been devastating the United States, and areas that experienced substantial middle class job loss also tended to be on the front lines of the opioid epidemic.³ If the changes we’ve seen so far pale in comparison with what might come, there is a real risk of future social and economic disruption on an unprecedented scale—as well as the rise of even more dangerous political demagogues who will thrive on the fear that is certain to accompany such a rapidly shifting landscape.

事实上，人工智能对经济的影响是一把双刃剑。一方面，它可能会提高生产力，使产品和服务更便宜，并推动能够改善我们生活的创新。人工智能有可能创造经济价值，这对于我们摆脱目前巨大的经济困境至关重要。另一方面，它几乎肯定会消除或淘汰数百万个工作岗位，同时将经济不平等推向更高的水平。除了失业和不断加剧的不平等带来的社会和政治影响外，还有另一个重要的经济后果：充满活力的市场经济依赖于大量能够购买所生产产品和服务的消费者。如果这些消费者没有工作，也没有收入，他们将如何创造推动经济持续增长所需的需求？

The reality is that artificial intelligence will be a dual-edged sword in terms of its economic impact. On one hand, it will likely increase productivity, make products and services more affordable and enable innovation that can improve all our lives. AI has the potential to create economic value that will be indispensable as we look toward digging ourselves out of the massive economic hole in which we now find ourselves. On the other hand, it is virtually certain to eliminate or deskill millions of jobs while driving economic inequality to even higher levels. Aside from the social and political implications of unemployment and ever-rising inequality, there is another important economic consequence: a vibrant market economy depends on vast numbers of consumers who are able to purchase the products and services being produced. If these consumers do not have jobs, and thus income, how will they create the demand necessary to drive continued economic growth?

人工智能与工作自动化：这次有什么不同吗？

AI AND JOB AUTOMATION: IS THIS TIME DIFFERENT?

人们担心机器有朝一日会取代工人并造成长期结构性失业，这种担忧由来已久，至少可以追溯到 200 多年前英国诺丁汉发生的卢德分子起义。此后几十年，人们一次又一次地敲响警钟。例如，在 20 世纪 50 年代和 60 年代，人们非常担心工业自动化会很快取代数百万个工厂工作岗位，从而导致大面积失业。然而，迄今为止，历史表明，经济总体上已经适应了不断进步的技术，创造了新的就业机会，而这些新工作往往需要更多的技能，工资也更高。

The fear that machines might someday displace workers and produce long-term, structural unemployment has a long history, stretching back, at a minimum, to the Luddite revolts that took place in Nottingham, England, over two hundred years ago. In the decades since, the alarm has been raised again and again. In the 1950s and 1960s, for example, there was a great deal of concern that industrial automation would soon displace millions of factory jobs, leading to widespread unemployment. So far, however, history shows that the economy has generally adjusted to advancing technology by creating new employment opportunities and that these new jobs often require more skills and pay higher wages.

历史上最极端的技术性失业案例之一（也是那些怀疑技术性失业是否会造成问题的人经常引用的案例研究）与美国农业机械化有关。19 世纪末，大约一半的美国工人从事农业。如今，这一数字在 1% 到 2% 之间。拖拉机、联合收割机和其他农业技术的出现不可逆转地蒸发了数百万个工作岗位。由于失业的农场工人迁移到城市寻找工厂工作，这一转变确实导致了严重的短期和中期失业。然而，最终，失业的工人被新兴的制造业所吸收，从长远来看，平均工资和整体繁荣大幅增加。后来，工厂实现自动化或迁往海外，工人们再次转型，这次是转向服务业。今天，近 80% 的美国劳动力受雇于服务业。

One of the most extreme historical examples of technologically-induced job losses—and a case study often cited by those who are skeptical that technological unemployment will ever pose a problem—concerns the mechanization of agriculture in the United States. In the late 1800s, about half of American workers were engaged in farming. Today, the number is between one and two percent. The advent of tractors, combine harvesters and other agricultural technology irreversibly vaporized millions of jobs. This transition did result in significant short- and medium-term unemployment as displaced farm workers migrated to cities in search of factory work. Eventually, however, the unemployed workers were absorbed by a rising manufacturing sector, and over the long run, average wages as well as overall prosperity increased dramatically. Later, factories automated or moved offshore, and workers transitioned again, this time to the service sector. Today, nearly eighty percent of the American workforce is employed in service industries.

关键问题是，人工智能对就业市场的冲击是否会导致类似的结果。人工智能是否只是另一种节省劳动力的创新，就像农业技术改变了农业一样？还是说它完全不同？我一直认为人工智能确实与众不同，其原因在于本书的核心论点：人工智能是一种系统的、通用的技术，与电力无异，因此它最终将扩展到我们经济和社会的各个方面。

The key question is whether the job market disruption that results from the impact of artificial intelligence will lead to a similar outcome. Is AI just another example of a labor-saving innovation like the agricultural technology that transformed farming? Or is it something fundamentally different? My argument has been that AI is indeed different, and the reason is anchored in the core thesis of this book: that artificial intelligence is a systemic, general-purpose technology not unlike electricity, and it will therefore ultimately scale across and invade every aspect of our economy and society.

从历史上看，劳动力市场的技术颠覆往往会对各个行业产生影响。农业机械化摧毁了数百万个工作岗位，但不断崛起的制造业最终可以吸收这些工人。同样，随着制造业实现自动化和工厂转移到低工资国家，快速增长的服务业为失业工人提供了机会。相比之下，人工智能将或多或少同时影响经济的每个部门。最重要的是，这将包括目前占美国劳动力绝大多数的服务业和白领工作。人工智能的触角最终将触及并改变几乎所有现有行业，未来出现的任何新行业都很可能从一开始就采用最新的人工智能和机器人创新。换句话说，似乎不太可能出现一个拥有数千万个新工作岗位的全新行业来吸收现有行业中被自动化取代的所有工人。相反，未来的行业将建立在数字技术、数据科学和人工智能的基础上——因此，它们根本不会创造大量的就业机会。

Historically, labor market technological disruptions have tended to impact on a sector-by-sector basis. Agricultural mechanization destroyed millions of jobs, but a rising manufacturing sector was available to eventually absorb those workers. Likewise, as manufacturing automated and factories offshored to low-wage countries, a rapidly growing service sector provided opportunities for the displaced workers. In contrast, artificial intelligence will impact every sector of the economy more or less simultaneously. Most importantly, this will include the service sector and white collar jobs that now engage the vast majority of the U.S. workforce. AI’s tentacles will eventually reach into and transform virtually every existing industry, and any new industries that arise in the future will very likely incorporate the latest AI and robotics innovations from their inception. In other words, it seems very unlikely that some entirely new sector with tens of millions of new jobs will somehow materialize to absorb all the workers displaced by automation in existing industries. Rather, future industries will be built on a foundation of digital technology, data science and artificial intelligence—and as a result, they will simply not generate large numbers of jobs.

第二点涉及工人所从事活动的性质。据合理估计，大约一半的劳动力从事的工作基本上是常规的、可预测的。4^我的意思不是“死记硬背”，而只是说这些工人往往会一次又一次地面对相同的基本任务和挑战。换句话说，工作的本质——或者至少是组成工作的很大一部分任务——基本上都包含在历史数据中，这些数据反映了工人在一段时间内所做的事情。这些数据最终将为机器学习算法提供丰富的资源，这些算法可以用来找出如何自动化其中许多任务。换句话说，我们面临的未来是，几乎所有类型的常规、可预测的工作最终都将消失，这对最适合这类工作的工人来说可能是一个特别困难的挑战。在整个二十世纪，不断进步的节省劳动力的技术促使工人转向不同的行业，但在大多数情况下，他们仍然从事着大量的常规工作。想象一下，从 1900 年的农场工人，到 1950 年的工厂流水线工人，再到今天沃尔玛扫描条形码的收银员。这些都是完全不同行业的不同工作，但它们都以基本常规和可预测的任务为定义。这一次，一些新行业不会有大量常规工作来容纳被取代的工人。相反，工人们将面临一种完全不同的转型，这种转型从根本上来说是非常规的，而且通常需要具备一些素质，例如有效地与他人建立关系的能力，或执行非常规的分析或创造性工作的能力。假设有足够数量的此类新工作，一些工人将成功完成这一转型，但许多其他人可能会挣扎。

A second point involves the nature of the activities undertaken by workers. It’s reasonable to estimate that roughly half our workforce is engaged in occupations that are largely routine and predictable in nature.⁴ By this, I don’t mean “rote-repetitive” but simply that these workers tend to face the same basic set of tasks and challenges again and again. In other words, the essence of the job—or at least a large fraction of the tasks that comprise it—is essentially encapsulated in historical data reflecting what the worker has done over time. Such data will eventually provide a rich resource for machine learning algorithms that can be turned loose to figure out how to automate many of these tasks. In other words, we’re facing a future where nearly all kinds of routine, predictable work are eventually going to evaporate, and this is likely to prove an especially difficult challenge for those workers who are best suited to such work. Throughout the twentieth century, advancing labor-saving technology drove workers to move to different sectors, but for the most part, they continued doing largely routine work. Imagine the transition from a farm worker in 1900, to a factory assembly line worker in 1950, to a cashier scanning barcodes at Walmart today. These are all very different jobs in entirely different sectors, but they are all defined by largely routine and predictable tasks. This time around, there aren’t going to be large numbers of routine jobs in some new sector to accommodate displaced workers. Instead, workers will be faced with making an entirely different kind of transition into work that is fundamentally non-routine and may often require qualities such as the ability to effectively build relationships with others or to perform non-routine analytical or creative work. Assuming that a sufficient number of such new jobs are available, some workers will successfully make this transition, but many others will likely struggle.

换句话说，我认为我们面临的情况是，我们劳动力的很大一部分最终面临被剥夺就业市场权利的风险。但有证据表明这样的事情真的发生了吗？毕竟，在冠状病毒大流行爆发之前，失业率远低于 4%。

In other words, I think we face a scenario where a significant fraction of our workforce is eventually at risk of being disenfranchised from the job market. But is there any evidence that anything like this has actually occurred? After all, the unemployment rate prior to the advent of the coronavirus pandemic was well below four percent.

冠状病毒大流行爆发前的故事

THE STORY UP UNTIL THE ONSET OF THE CORONAVIRUS PANDEMIC

从 2009 年大衰退结束到 2020 年 1 月的十年间，这是有史以来最长的战后经济复苏期，失业率从 10% 下降到 3.6%，低于过去五十年的任何记录。5^但需要注意的是，这一总体失业率是根据美国人口普查局进行的家庭调查测算出来的，只包括那些积极寻找工作的工人。任何想找工作但灰心丧气并放弃的人，或者认为没有他们愿意接受的工作的人，都不被算作失业者。

Over the ten-year period from the end of the Great Recession in 2009 until January 2020, the longest postwar economic recovery on record, the unemployment rate fell from ten percent to 3.6 percent—a level lower than any recorded in the past fifty years.⁵ An important caveat, however, is that this headline unemployment rate, which is measured on the basis of a household survey conducted by the U.S. Census Bureau, captures only those workers who are actively seeking employment. Anyone who would like to have a job but has become disheartened and given up or who believes there are no jobs available that they would be willing to accept, is not counted as unemployed.

要了解完全脱离劳动力队伍的人数，查看劳动力参与率很有用。这方面的情况远不如总体失业率乐观。

To get some insight into the number of people who have become completely detached from the labor force, it’s useful to look at the labor force participation rate. The story here is far less positive than the headline unemployment rate.

如图1所示，在职或积极寻找工作的黄金年龄男性的比例从 1965 年的 97% 左右下降到 2014 年的最低点 88%，然后在 2020 年 1 月略有回升至约 89%。6 在^此期间，完全被剥夺就业权利的男性数量几乎增加了三倍。退出就业市场的男性似乎会选择社会保障残疾计划，该计划在 2007 年至 2010 年间申请人数激增。7^鉴于没有证据表明工作场所伤害泛滥，该计划很可能被那些在就业市场机会不多的工人作为最后的收入来源。虽然对男性劳动力参与率的影响最为显著，但总体统计数据显示，自本世纪初的二十年来，情况大致相似。

As Figure 1 shows, the percentage of prime age working men in work or actively seeking employment has fallen from about ninety-seven percent in 1965 to a nadir of eighty-eight percent in 2014 before recovering slightly to about eighty-nine percent in January 2020.⁶ The number of men entirely disenfranchised from the job market has nearly quadrupled over this time. One destination for men who have been exiting the job market appears to be the Social Security Disability program, which saw a surge of applications between 2007 and 2010.⁷ Given that there was no evidence of an epidemic of workplace injuries, it seems likely that the program is being used as an income of last resort by workers who see few viable job market opportunities. While the impact on labor force participation for men has been the most dramatic, the overall statistics show a broadly similar story over the two decades since the turn of the century.

图 2显示了 18 至 64 岁所有工人（包括男性和女性）的劳动力参与率。8 ²⁰⁰⁰ 年之前参与率的上升反映了更多女性进入劳动力市场。然而，在那次高峰之后，随着男性和女性都退出劳动力市场，这一趋势一直呈下降趋势。换句话说，即使失业率降至历史最低水平，也有越来越多的完全被剥夺权利的工人，尽管整体叙事指向就业市场蓬勃发展，但他们仍然基本不为人知。虽然技术变革肯定不是这里唯一的因素，但工厂和办公室中高薪常规工作的不断自动化可能发挥了重要作用。

Figure 2 shows the workforce participation rate for all workers aged eighteen to sixty-four, including both men and women.⁸ The rising participation rate up to the year 2000 reflects the entry of more women into the workforce. After that peak, however, the trend has been downward as both men and women have exited the labor market. In other words, even as the unemployment rate fell to historic lows, there has been an ever-increasing mass of completely disenfranchised workers who remained largely invisible as the overall narrative pointed to a booming job market. While technological change has certainly not been the only factor at play here, the relentless automation of well-paid routine jobs in factories and offices has likely played a significant role.

第二个重要趋势是生产力与工资脱钩，以及不断加剧不平等。劳动生产率是衡量工人效率的指标，等于总经济产出除以生产该产出所需的劳动小时数。生产力也许是所有经济指标中最重要的。高生产率是区分富裕发达国家和贫穷国家的决定性特征。随着工作场所采用的技术进步，以及工人教育和健康等其他因素的改善，工人可以生产更多。因此，他们应该能够要求更高的工资，因此，提高生产力实际上会将钱存入几乎所有工人的口袋，是推动国家全面繁荣的关键因素。至少这是标准的经济叙述。

A second important trend involves the decoupling of productivity and wages, along with a relentless drive toward increased inequality. Labor productivity is a measure of worker effectiveness and is equal to total economic output divided by the number of labor hours required to generate that output. Productivity is perhaps the most important of all economic metrics. A high rate of productivity is a defining characteristic that distinguishes a wealthy, developed nation from an impoverished one. As the technology employed in workplaces advances, and as other factors such as worker education and health likewise improve, workers can produce more. As a result, they should be able to command higher wages, and therefore rising productivity essentially deposits money into the pockets of nearly all workers and is a critical driver of broad-based national prosperity. At least that is the standard economic narrative.

图 1. 25 至 54 岁男性劳动力参与率

Figure 1. Workforce Participation Rate, Men Ages Twenty-Five to Fifty-Four

图 2. 总体劳动力参与率

Figure 2. Overall Workforce Participation Rate

然而，如图 3所示，至少自 20 世纪 70 年代以来，工人的报酬一直未能跟上生产率的提高，两条线之间的鸿沟越来越大。9其结果是^，技术进步和生产率提高带来的几乎所有收益都被收入分配顶层附近的一小群人所获取。换句话说，企业主、经理、超级明星员工和投资者正在获取进步的果实，而普通工人几乎一无所获。值得注意的是，该图反映了商业部门所有工人的报酬，其中包括高层管理人员、超级明星运动员和演艺人员，以及其他高薪工人。如果该图仅反映占美国劳动力约 80% 的普通非管理人员，生产率和报酬之间的差距会更大。

As Figure 3 shows, however, at least since the 1970s, compensation for workers has failed to track rising productivity and an ever-widening chasm has opened between the two lines.⁹ The upshot of this is that nearly all the gains from technological progress and improving productivity are now being captured by a relatively small group of people near the top of the income distribution. In other words, business owners, managers, superstar employees and investors are capturing the fruits of progress, and ordinary workers are getting almost nothing. It’s worth noting that this graph reflects compensation for all workers in the business sector, and that includes top-level executives, superstar athletes and entertainers, as well as other highly paid workers. If the graph instead reflected only the average non-supervisory workers who account for about eighty percent of the U.S. workforce, the gap between productivity and compensation would be even greater.

我认为，这两条线之间的差距不断扩大，至少在一定程度上是由工作场所部署的机器和技术的变化性质所致。在二战后美国的“黄金时代”，图上的两条线紧密相连，工作场所使用的机器显然是工人操作的工具；随着工具的改进，工人的产出增加，他们的价值也随之增加。然而，随着技术在几十年后的不断进步，工作场所使用的许多机器逐渐变得更加自动化，技术越来越多地替代而不是补充劳动力。换句话说，技术现在正在使越来越多的工人价值降低而不是提高。这反过来又使工人更容易被替代，降低了他们的议价能力，并在生产力不断提高的同时压低了薪酬。

I would argue that the widening divergence between these two lines is driven, at least in part, by the changing nature of the machines and technology deployed in workplaces. During America’s “golden age” following World War II, the two lines on the graph were tightly coupled and the machines used in workplaces were clearly tools operated by workers; as the tools improved, the workers’ output increased and they became more valuable. As technology has continued to advance in the decades since, however, many of the machines used in workplaces are gradually becoming more autonomous, and technology is increasingly substituting for, rather than complementing, labor. In other words, technology is now making an increasing fraction of workers less rather than more valuable. This, in turn, makes workers more interchangeable, reduces their bargaining power, and acts to push down compensation even as productivity continues to increase.

生产力与报酬的脱钩直接导致收入不平等加剧。随着技术取代或降低劳动力的价值，资本占据了企业利润的更大份额。过去二十年，美国以及其他许多发达国家都出现了劳动力在国民收入中所占份额下降的情况。由于资本所有权高度集中在富人手中，收入从劳动力转移到资本就相当于从多数人到少数人的重新分配，这加剧了收入不平等。在美国，这种趋势尤其引人注目，基尼系数的上升生动地证明了这一点。该指数是衡量财富集中程度的指标。在极端情况下，基尼值为零表示一个国家中每个人都拥有平等的财富份额，而值为 100 则表示一个人拥有整个国家的财富。实际值一般在 20 到 50 之间，数字越高表示不平等程度越高。在美国，基尼系数从 1986 年的 37.5 上升至 2016 年的 41.4，创历史新高^。10

This decoupling of productivity and compensation leads directly to increased income inequality. As technology displaces or diminishes the value of labor, a larger share of business profits is captured by capital. This decline in labor’s share of national income has been found over the last two decades in the United States as well as in a variety of other developed countries. Because capital ownership is highly concentrated in the hands of the wealthy, a redirection of income from labor to capital amounts to a redistribution from the many to the few, and this increases income inequality. In the United States the trend has been especially dramatic and is demonstrated vividly by the rise of the Gini coefficient. This index is a measure of the concentration of wealth. At the extremes, a Gini value of zero would indicate that everyone in a country has an equal share of wealth, and a value of 100 would mean that a single individual owns all the nation’s wealth. Realistic values generally fall between roughly 20 and 50, with a higher number indicating more inequality. In the U.S., the Gini coefficient rose from 37.5 in 1986 to 41.4 in 2016—a level higher than any previously recorded.¹⁰

图 3.生产力与薪酬

Figure 3. Productivity vs. Compensation

收入不平等加剧的趋势部分是由于美国提供的工作质量普遍下降。近几十年来，美国就业岗位越来越多地集中在服务业的低薪工作上。这些工作涉及零售销售、食品准备和服务、保安、办公室和酒店的清洁或清洁工作，收入微薄，福利很少，而且往往不是全职工作，工作时间也不稳定。零工经济的兴起进一步加剧了这一趋势，在零工经济中，工人根据任务完成情况获得报酬，几乎没有可预测的收入保证，也很少或根本无法获得其他工人享有的法律保障。布鲁金斯学会 2019 年 11 月的一份报告发现，美国劳动力中有 44% 从事低薪工作，平均年收入约为 18,000 美元^。11

This trajectory toward rising income inequality has been driven in part by a general decline in the quality of jobs on offer in the United States. In recent decades, American job creation has been weighted increasingly toward low-wage jobs in the service sector. These jobs, in areas like retail sales, food preparation and serving, security and cleaning or janitorial jobs in offices and hotels, provide minimal incomes and few if any benefits, and are often less than full-time with unreliable hours. The rise of the gig economy, in which workers receive payments based on a task-completion basis with virtually no guarantee of a predictable income and little or no access to the legal safeguards provided to other workers, has further exacerbated the trend. A November 2019 report from the Brookings Institution found that a full forty-four percent of the U.S. workforce is engaged in low-wage jobs providing income averaging about $18,000 per year.¹¹

2019 年，一组研究人员开发了一项新的经济指标，这尤其明显地改变了美国工人可从事工作性质的变化。美国私营部门就业质量指数衡量的是好工作（收入高于平均水平的工作）与低质量工作（收入低于平均水平的工作）的比例。12^指数值为 ¹⁰⁰表示好工作和低质量工作数量相等，而指数低于 100 则表示低质量工作占据了就业市场主导地位。从 1990 年到 2019 年底的 30 年间，该指数从 95 暴跌至 81。13这种质量下降可能与工厂和办公室等环境中大部分常规但高薪工作的消失密切相关。这些工作曾经是美国中产阶级的支柱，但却遭到技术和全球化的无情摧毁。

This change in the nature of the jobs available to American workers was made especially evident when a group of researchers developed a new economic metric in 2019. The U.S. Private Sector Job Quality Index measures the ratio of good jobs, defined as those that provide an above-average income, to low-quality jobs, which offer an income below the average.¹² An index value of 100 indicates an equal number of good- and low-quality jobs, while a value below 100 indicates that lower-quality jobs dominate the employment landscape. Over the 30 years between 1990 and the end of 2019, the index plunged from 95 to 81.¹³ This decline in quality is likely closely tied to the evaporation of largely routine but well-paying jobs in environments like factories and offices. These are the jobs that once formed the backbone of the American middle class but have been relentlessly destroyed by both technology and globalization.

当然，经济也创造了更高技能、更高薪的工作，但对于近四分之三没有四年制大学学位的美国工人来说，这些工作很少能获得。即使在大学毕业生中，就业不足也是一个严重且日益严重的问题。大学毕业生在担任咖啡师或快餐店员工的同时，背负着巨额学生贷款债务的故事屡见不鲜。纽约联邦储备银行 2020 年 2 月发布的数据显示，41% 的应届大学毕业生从事的工作不需要大学学位。对于整个大学毕业生来说，就业不足的数字是三分之一。尽管整个经济的总体失业率下降到 3.6%，但 22 至 27 岁的应届大学毕业生的失业率却超过 6%。14换句话说，尽管传统观点认为我们需要更加重视教育并扩大大学入学率，但经济根本无法创造足够的技术就业机会^来吸收已经毕业的毕业生。

The economy has, of course, also created higher-skill, higher-paying jobs, but these are rarely accessible to the nearly three quarters of American workers who lack a four-year college degree. And even among college graduates, underemployment is a serious and growing problem. Stories of college graduates weighed down by massive student loan obligations while working as baristas or fast food workers are all too common. Data published by the Federal Reserve Bank of New York in February 2020 showed that a full forty-one percent of recent college graduates are engaged in jobs that do not require a college degree. For college graduates as a whole, the underemployment figure is one in three. And recent college graduates aged twenty-two to twenty-seven had an unemployment rate of more than six percent even as headline unemployment across the economy as a whole fell to 3.6 percent.¹⁴ In other words, even as the conventional wisdom suggests that we need to put more emphasis on education and expand college enrollment, the economy is simply not creating enough skilled job opportunities to absorb the graduates already being produced.

收入不平等加剧和就业质量下降不仅对直接受影响的个人来说是坏消息。相反，它们还破坏了推动我们实现持续经济活力所需的市场需求。美国经济约 70% 与个人消费支出直接相关。然而，即使是这一比例也低估了消费需求的重要性，因为商业投资也与消费需求息息相关。例如，想想波音生产的飞机（当然不是消费品）是如何被航空公司购买的，只有它们反过来预测到消费者对机票的需求。当然，这种经济依赖性因冠状病毒危机的影响而凸显出来。

The rise in income inequality and the decline in job quality are not just bad news for the individuals directly impacted. Rather, they undermine the market demand required to drive us toward sustained economic vitality. Roughly seventy percent of the U.S. economy is associated directly with individual consumer spending. Even that fraction, however, underestimates the importance of consumer demand because business investment is also tied to consumer demand. For example, think about how airplanes produced by Boeing—certainly not a consumer item—are purchased by airlines only if they, in turn, anticipate consumer demand for plane tickets. This economic dependency has, of course, been brought into stark relief by the impact of the coronavirus crisis.

就业是将购买力交付给消费者的主要机制。随着收入分配变得更加不平等，大多数工人，也就是消费者，可支配的收入越来越少。在过去的几十年里，少数富人的收入大幅增加，但这一小部分人根本无法也不愿消费到足以弥补收入分配较低水平的可支配收入损失的程度。换句话说，对经济增长至关重要的对产品和服务的广泛消费者需求正在逐渐减少。

Jobs are the primary mechanism that delivers purchasing power into the hands of consumers. As the distribution of income becomes more unequal, the bulk of workers, and therefore consumers, are left with less discretionary income. Over the past few decades, income has risen dramatically for the wealthy few, but this small fraction of the population simply cannot and will not spend to a degree that makes up for the loss of discretionary income at lower levels of the income distribution. In other words, the broad-based consumer demand for products and services that is vital to generating economic growth is gradually eroding.

消费需求低迷的证据是失业率和通货膨胀之间正常关系的破裂。1958 年，经济学家威廉·菲利普斯 (William Phillips) 指出，失业率和通货膨胀率之间通常存在着一致的权衡。失业率下降，通货膨胀率上升。我在大学学习经济学时，这种反比关系被称为菲利普斯曲线，是该领域的基本原理之一。然而，自 2009 年大衰退结束以来的几年里，这种关系已经破裂，低失业率现在与极低的通货膨胀率和低利率并存。15^我认为，一个重要原因是失业率下降不再与足以推动通货膨胀的工资或消费者需求增长有关。随着技术进步和全球化削弱了大多数普通工人讨价还价提高工资的能力，将购买力交到消费者手中并推动需求增长的机制变得越来越无效。

Evidence for underwhelming consumer demand has manifested in the breakdown in the normal relationship between unemployment and inflation. In 1958, the economist William Phillips showed that there is generally a consistent trade-off between unemployment and inflation. As unemployment falls, inflation rises. When I studied economics in college, this inverse relationship, known as the Phillips curve, was taught as one of the basic principles of the field. In the years since the end of the Great Recession in 2009, however, this relationship has broken down and low unemployment now coexists with very low rates of inflation and low interest rates.¹⁵ I believe that an important reason for this is that falling unemployment is no longer associated with increases in wages or consumer demand that are sufficient to drive inflation. As advancing technology, along with globalization, has eroded the ability of most average workers to bargain for higher wages, the mechanism that gets purchasing power into the hands of consumers and drives increasing demand has become less and less effective.

另一个证据是，美国大型企业一直持有大量现金，其中大部分投资于利率处于历史低位的美国国债。截至 2018 年底，美国企业持有约 2.7 万亿美元现金。16如果这些公司的高管看到了商品和服务需求旺盛的证据，他们为什么不把更多的钱投资于新产品的开发或提高产量以满足日益增长的需求呢？在没有强劲需求的情况下，美国经济只能^实现中等增长，并且依赖于美联储将利率维持在异常低的水平，即使失业率降至 4% 以下。

More evidence comes from the fact that large U.S. corporations have been sitting on enormous amounts of cash, much of which is invested in U.S. Treasury securities paying historically low interest rates. As of the end of 2018, American businesses were hording about $2.7 trillion.¹⁶ If the executives running these companies saw evidence of vibrant demand for goods and services, why wouldn’t they invest more of this money in developing new products or ramping up production to meet that increasing demand? Without robust demand, the U.S. economy has managed only middling rates of growth and has become dependent on the Federal Reserve holding interest rates at unusually low levels even as the unemployment rate fell below four percent.

消费需求不温不火的另一个重要影响是，它削弱了生产力增长。对人工智能和机器人对就业市场的影响持怀疑态度的经济学家很快指出，如果机器确实在快速取代劳动力，那么随着剩余工人生产出越来越多的产品，我们应该看到劳动生产率飙升。在生产率没有飙升的情况下，经济学家们对机器人抢走工作岗位的担忧不以为然。这一断言的问题在于，产出完全取决于需求。除非有客户愿意购买该产品，否则任何企业都不会继续生产商品或服务。（我在 2015 年出版的《 机器人崛起》一书中也更详细地解释了生产力受需求限制的观点。我发现经济学家们没有更多地关注这个问题，而是倾向于简单地宣称缺乏“生产力飙升”证明工作自动化不是问题，这让我感到有些惊讶^。17）

Another important implication of tepid consumer demand is that it undermines productivity growth. Economists who are skeptical of the impact of artificial intelligence and robotics on the job market are quick to point out that if machines were indeed substituting for labor at a rapid clip, we should see soaring labor productivity as the remaining workers produced ever more output. In the absence of skyrocketing productivity, the economists brush aside any concerns about robots stealing jobs. The problem with this assertion is that output is entirely dependent on demand. No business will continue to produce goods or services unless there is a customer prepared to purchase that output. (I also explained the idea that productivity is limited by demand in more detail in my 2015 book Rise of the Robots. I find it somewhat surprising that economists don’t focus more on this issue and instead tend to simply declare that a lack of “soaring productivity” proves that job automation is a non-issue.¹⁷)

假设有一名理发师，他的工作效率可以用每小时理发的次数来衡量。很多因素都会影响效率。这名理发师是否受过良好的培训，拥有优质的工具？是否有稳定的电力供应来维持设备运转？这些都是经济学家关注的重点。但还有一点也至关重要：前来理发的顾客数量。如果理发师排着长队，效率就会很高。如果只有偶尔一位顾客进来，效率就会很低——无论培训或理发技术有多好。

Imagine a worker whose job is to give haircuts. The productivity of that worker might be measured in terms of the number of haircuts performed per hour. Lots of things would affect that rate of productivity. Does the worker have good training and quality tools? Is there a stable supply of electricity to keep the equipment running? These are the kinds of things that economists tend to focus on. But there is something else that is absolutely critical: the number of customers who show up for a haircut. If there is a long line of eager customers, productivity will be high. If only an occasional client wanders in, productivity will be low—regardless of how good the training or haircutting technology is.

当我与麦肯锡全球研究所 (MGI) 主席詹姆斯·曼尼卡交谈时，我提出了“生产率增长受需求限制”这一观点。麦肯锡全球研究所开展了许多重要研究，重点关注技术对企业和经济的影响。正如曼尼卡所解释的那样：

This idea that productivity growth is limited by demand came up when I spoke with James Manyika, the chairman of the McKinsey Global Institute (MGI), which has conducted numerous important studies focused on the impact of technology on businesses and the economy. As Manyika explained:

我们也知道需求的关键作用——大多数经济学家，包括麦肯锡全球研究院的经济学家，经常关注生产力的供应侧效应，而较少关注需求侧。我们知道，当需求大幅放缓时，即使生产效率再高，测算出来的生产力也不会很好。这是因为生产力测量有一个分子和一个分母：分子涉及增值产出的增长，这要求产出被需求吸收。因此，如果需求因某种原因而滞后，就会损害产出增长，从而降低生产力增长，无论技术进步如何^。18

We also know the critical role of demand—most economists, including here at MGI, have often looked at the supply-side effects of productivity, and not as much at the demand side. We know that when you’ve got a huge slowdown in demand you can be as efficient as you want in production, and measured productivity still won’t be great. That’s because the productivity measurement has a numerator and a denominator: the numerator involves growth in value-added output, which requires that output is being soaked up by demand. So, if demand is lagging for whatever reason, that hurts growth in output, which brings down productivity growth, regardless of what technological advances there may have been.¹⁸

归根结底，在新冠疫情爆发之前的几年里，美国经济就像一辆崭新闪亮的汽车，但引擎盖下却存在着严重的问题。失业率看起来不错，但越来越多的人口被完全抛在后面。不平等现象急剧上升，大多数工人不再因技术进步而获得更多的繁荣。随着不平等现象变得越来越严重，推动消费需求的收入分配机制正在逐渐瓦解，这反过来又破坏了经济增长，抑制了对未来繁荣至关重要的持续生产力增长。这场疫情彻底颠覆了一切，使我们陷入了前所未有的经济危机，但所有这些趋势仍然存在，并可能产生阻力，使我们从目前的困境中恢复过来变得更加困难。

The bottom line is that the years leading up to the onset of the coronavirus pandemic delivered an American economy that was a bit like a shiny, newly painted car—but with serious problems under the hood. The unemployment rate looked great, but a large and growing fraction of the population was being left completely behind. Inequality has risen dramatically, and most workers are no longer experiencing increased prosperity as a result of technological progress. And as things have become ever more unequal, the mechanism that distributes the income that powers consumer demand is eroding, and that in turn is undermining economic growth and dampening down the sustained rise in productivity that is crucial to future prosperity. The pandemic has completely upended things and plunged us into an unprecedented economic crisis, but all these trends remain in place and will likely produce headwinds that will make it even more challenging to recover from our current predicament.

新冠肺炎疫情后及复苏

POST COVID-19 AND RECOVERY

新冠病毒大流行引发了一场史无前例的全球经济危机。在美国和世界各国，数百万人几乎在一夜之间失去工作，整个行业几乎停摆，经济陷入了自 1930 年代大萧条以来最严重的衰退。截至 2020 年 12 月，失业率接近 7%，所有迹象都表明，在疫苗大规模部署开始在 2021 年中期左右扭转局面之前，情况可能会变得更糟。美国对疫情的管理不力导致病毒大规模复发，截至 2021 年 1 月，该国单日新增新冠肺炎死亡人数超过 4,000 人。随着住院人数激增，美国各州和地方政府再次强制企业关闭，而在英国和许多欧洲国家，全国封锁再次成为常态。换句话说，即使至少有两种有效疫苗开始分发，危机的经济影响似乎仍将持续一段时间。

The coronavirus pandemic has unleashed a global economic crisis of unprecedented ferocity. In the United States and in countries throughout the world, millions of jobs have been lost very nearly overnight, entire sectors have been virtually shut down and the economy has plunged into the deepest downturn since the Great Depression in the 1930s. As of December 2020, the unemployment rate stands at nearly seven percent, and all indications are that things could well get worse before wide deployment of vaccines begins to bend the curve sometime around the middle of 2021. The U.S.’s bungled management of the pandemic led to a widespread resurgence of the virus, and the country is recording more than 4,000 COVID-19 deaths in a single day as of January 2021. As hospitalizations surge, U.S. state and local governments are once again forcing businesses to close, while in the United Kingdom and many European countries national lockdowns are again the rule. In other words, even as at least two effective vaccines are starting to be distributed, the economic impact of the crisis seems poised to linger for some time to come.

现实情况是，由于自动化和更广泛的技术影响，所有这些都为就业市场的急剧转变创造了肥沃的土壤。历史表明，由于采用节省劳动力的技术而导致的绝大多数失业往往集中在经济低迷时期。常规工作受到的打击尤其严重，这在很大程度上解释了稳定的中产阶级工作岗位的消失，以及它们最终被服务业中不太理想、工资较低的工作机会所取代。事实上，经济学家 Nir Jaimovich 和 Henry E. Siu 研究了这一现象，并在 2018 年的一篇论文中发现，“几乎所有常规职业的就业损失都发生在经济低迷时期。” ¹⁹ 似乎正在发生的事情是，企业在经济压力下裁员；然后随着经济衰退的加剧，他们采用新技术并重组工作场所；最终，一旦经济复苏，他们就会发现他们能够避免重新雇用他们之前认为对其运营至关重要的所有或大部分工人。当前经济衰退的深度表明，大多数企业将面临提高效率的巨大压力，危机持续的时间越长，他们就越有时间将新技术（包括人工智能的最新应用）融入其商业模式。

The reality is that all this creates a fertile ground for a dramatic transformation of job markets due to automation and the impact of technology more broadly. History shows that the vast majority of job losses from the adoption of labor-saving technology tend to be concentrated in economic downturns. Routine jobs are especially hard hit, and this largely explains the evaporation of solid middle class jobs and their eventual replacement by less desirable, lower-wage opportunities in the service sector. Indeed, the economists Nir Jaimovich and Henry E. Siu studied this phenomenon and in a 2018 paper found that “essentially all employment loss in routine occupations occurs in economic downturns.”¹⁹ What appears to be happening is that businesses eliminate workers under economic duress; then as the downturn progresses, they incorporate new technology and reorganize workplaces; eventually once recovery occurs, they find that they’re able to avoid re-hiring all or most of the workers they previously believed to be essential to their operations. The depth of the current downturn suggests that most businesses will be under enormous pressure to become more efficient, and the longer the crisis lasts, the more time they will have to assimilate new technology—including the latest applications of artificial intelligence—into their business models.

除了采用新技术的纯粹经济动力之外，当前的危机的独特之处在于，它又增加了向更自动化工作场所过渡的另一个动机。正如我们在第 3 章中看到的那样，保持社交距离的需求已经大大推动了机器人技术在各个领域的采用。例如，美国和其他地方的肉类加工厂在数百或数千名工人几乎肩并肩工作的环境中，一次又一次成为主要的传染源。在这样的环境中，采用更多的自动化来降低工人密度可能是不可避免的。20虽然这是一个极端的情况，但几乎所有其他类型的工作环境，从工厂和仓库^到零售店再到办公室，情况也是如此。用机器人或智能算法取代工人，直接意味着近距离接触的人更少。面向客户的服务企业可能会认为，将直接的人际互动降到最低是一种营销优势，而就在几个月前，这种互动还被视为积极而非消极的。事实上，这种趋势已经显现：2020 年 7 月，快餐连锁店 White Castle 宣布将开始部署汉堡烹饪机器人，以“减少烹饪过程中人类与食物的接触，从而降低食物病原体传播的可能性”。21^这些因素的长期影响可能在一定程度上取决于危机的持续时间。然而，截至撰写本文时，这种情况似乎可能会持续很长时间，以至于至少部分因疫情而出现的行为变化和客户偏好将根深蒂固，甚至可能成为永久性的。

Beyond the purely economic impetus to adopt new technology, the current crisis is unique in that it adds yet another incentive to transition to a more automated workplace. As we saw in Chapter 3, the need for social distancing has already driven a dramatic boost to the adoption of robotic technologies in a variety of areas. Meat packing plants in the United States and elsewhere, for example, have again and again become major nexuses of contagion in settings where hundreds or thousands of workers labor virtually shoulder to shoulder. In environments like this, it’s likely inevitable that more automation will be adopted as a way of reducing worker density.²⁰ While this represents an extreme case, the same is true in virtually every other type of work environment, from factories and warehouses to retail stores to offices. Replacing workers with robots or with smart algorithms translates directly to fewer people in close proximity. Customer-facing service businesses are likely to perceive a marketing advantage in minimizing the direct human interaction that just a few months ago was viewed as a positive rather than a negative. Indeed, this trend is already in play: in July 2020, the fast food chain White Castle announced that it would begin deploying hamburger cooking robots in order to create “an avenue for reduced human contact with food during the cooking process—reducing potential for transmission of food pathogens.”²¹ The long-term impact of these factors will probably depend to some extent on the duration of the crisis. As of this writing, however, it seems likely that the situation will persist long enough that at least some of the behavioral changes and customer preferences that have emerged as a result of the pandemic will become ingrained and possibly permanent.

人工智能对职场的影响不会是机器人抢走工作这种简单的说法。研究表明，在大多数情况下，新技术的部署与现有工作之间并没有一一对应关系。相反，最容易被自动化取代的往往是具体任务，而不是整个职业。麦肯锡全球研究院 2017 年的一项有影响力的分析发现，目前全球工人执行的所有任务中，大约一半理论上已经可以通过现有技术实现自动化。麦肯锡的分析显示，只有 5% 的工作面临完全自动化的直接风险，但“在约 60% 的职业中，至少三分之一的组成活动可以实现自动化，这意味着所有工人的职场都将发生重大转变和变化。” ²² 很容易看出，如果两三名工人执行的任务中有相当一部分可以实现自动化，那么就有明显的潜力重新定义工作之间的界限并整合剩余的工作。经济压力以及降低工作场所密度的需求很可能将强烈激励许多组织重新思考和组织其工作环境，以利用这些未实现的效率，而这种趋势将因结合深度学习最新进展的更强大的应用程序的出现而进一步放大。在大多数情况下，结果将是工作岗位减少，而这些工作岗位很可能由具有完全不同技能和才能的不同员工担任。

The impact of artificial intelligence on workplaces is not going to be a straightforward narrative of robots stealing jobs. Research has shown that in most cases there is not a one-to-one correspondence between the new technology deployed and an existing job. Rather, it tends to be specific tasks—not entire occupations—that are most susceptible to automation. An influential 2017 analysis by the McKinsey Global Institute found that roughly half of all tasks currently performed by workers globally could, in theory, already be automated using existing technology. McKinsey’s analysis showed that only five percent of jobs were at immediate risk of full automation, but that “in about 60 percent of occupations, at least one-third of the constituent activities could be automated, implying substantial workplace transformations and changes for all workers.”²² It’s easy to see that if a significant fraction of the tasks being performed by two or three workers can be automated, then there is clear potential for redefining the boundaries between jobs and consolidating the remaining work. It seems very likely that economic pressure along with the need for reduced density in workplaces will create a powerful incentive for many organizations to rethink and reorganize their work environments in order to take advantage of these unrealized efficiencies, and this trend will be amplified by the arrival of far more capable applications incorporating the latest advances in deep learning. The result, in most cases, will be fewer jobs—and those jobs may well be held by different workers with entirely different skillsets and talents.

除了工作和任务的直接自动化之外，第二个重要力量是工作技能的降低。换句话说，新技术的采用使得曾经需要大量技能和经验的职位可以由几乎没有受过培训的低薪工人或零工经济中可互换的独立承包商来担任。一个典型的例子是伦敦著名的“黑色出租车”司机的经历。传统上，要获得驾驶这种出租车的执照，需要完全记住城市中几乎所有的街道，这是一个被称为获取“知识”的艰苦过程。所需的记忆范围非常广泛，以至于伦敦大学学院神经学家埃莉诺·马奎尔 (Eleanor Maguire) 的一项分析发现，与其他职业的工人相比，黑色出租车司机的海马体（与长期记忆相关的大脑区域）平均更大。23 要求未来司机获得“知识”这一要求在历史上为进入该行业提供了一道令人生畏的门槛，从而确保了出租车司机获得稳定的中产阶级工资。^随着 GPS 和智能手机导航应用程序的出现，这种情况发生了巨大变化。现在，司机即使对伦敦的街道一无所知，但只要有智能手机，就可以直接竞争，而拼车服务和其他类似出租车的选择对伦敦出租车司机的生计产生了巨大的负面影响。总的来说，去技能化会压低工资，因为这样可以让几乎没有或根本没有培训或经验的人也能从事这项工作，同时使工人更容易互换。这反过来又让企业能够容忍高流动率，进一步削弱工人的议价能力。随着自动化和去技能化的进步，我们有充分的理由预计不平等现象会加剧，创新的成果将继续越来越多地流向收入分配的顶层。

Aside from the direct automation of jobs and tasks, a second important force is the de-skilling of jobs. In other words, the adoption of new technology allows a role that once required significant skill and experience to instead be filled by a lower-wage worker with little training, or by an interchangeable independent contractor working in the gig economy. A classic example of this is the experience of the famous “black cab” taxi drivers in London. Obtaining a license to drive such a taxi traditionally requires full memorization of virtually all the streets in the city, a laborious process known as acquiring “The Knowledge.” The memorization required is so extensive that an analysis by the University College London neuroscientist Eleanor Maguire found that the hippocampus—the area of the brain associated with long-term memory—is, on average, larger in black cab drivers as compared to workers in other occupations.²³ This requirement for prospective drivers to acquire The Knowledge has historically provided a forbidding entry barrier into the profession and thereby ensured cab drivers a solid middle class wage. This has changed dramatically with the advent of GPS and smartphone navigation apps. Now drivers with no knowledge of London streets whatsoever, but access to a smartphone, are able to compete directly, and the onslaught from ride-sharing services and other taxi-like options has had a dramatic and negative impact on the livelihoods of London taxi drivers. In general, de-skilling acts to push down wages by making the job accessible to people will little or no training or experience, while at the same time making workers more interchangeable. This in turn allows businesses to tolerate high turnover rates and further undermines the bargaining power of workers. As both automation and de-skilling progress, there is every reason to expect that inequality will grow and that the fruits of innovation will continue to accrue increasingly to the top of the income distribution.

这些技术趋势将与疫情的其他重要影响交织在一起。例如，白领阶层全面采用远程办公方式，摧毁了办公楼集中区周围的商业生态圈。这种向远程办公的转变很可能至少在某种程度上是永久性的。例如，Facebook 已经宣布，其许多员工将能够无限期地选择远程办公。24在^这些曾经人头攒动的商业区，餐馆、酒吧和其他为办公室工作人员提供服务的企业的就业岗位可能永远不会恢复到以前的水平。清洁和维护办公室并提供安全保障的服务人员的就业岗位也可能受到影响。第二个关键因素是，提供这些工作岗位的大量小企业可能会破产。据一些统计，在疫情期间被迫关闭的小企业中，多达一半可能永远不会重新开业。25^最终，这些小企业曾经占据的市场份额将被更大、更具弹性的零售和餐饮连锁店所占据。然而，由于这些大型企业拥有更多的财务资源和内部专业知识，它们将更有能力成为新劳动力节省技术的早期采用者。换句话说，大型企业日益占据市场主导地位可能会直接加速服务业的工作自动化和技能降低。所有这些力量的汇聚很可能会对低薪服务业的再生产生重大抑制作用，而这些工作近年来一直是美国创造就业机会的主要引擎，这可能会使当前危机后的持续复苏变得更加困难。

These technological trends will intertwine with other important ramifications of the pandemic. For example, the wholesale adoption of remote work among white collar workers has decimated the business ecospheres that surround concentrations of office buildings. It seems very likely that the shift toward telecommuting will, at least to some extent, be permanent. Facebook, for example, has already announced that many of its employees will be able to opt for remote work indefinitely.²⁴ In these once teaming business districts, jobs at restaurants, bars and other businesses that cater to office workers may never return to previous levels. Jobs for the service workers who clean and maintain offices and provide security could likewise be impacted. A second key factor is the likely bankruptcy of a large fraction of the small businesses that disproportionately provide these jobs. By some accounts, up to half of the small businesses that were forced to shut down amidst the pandemic may never reopen.²⁵ Eventually, the market share once commanded by these small businesses will be captured by larger, more resilient retail and restaurant chains. However, because these larger businesses have greater financial resources and internal expertise, they will be far better positioned to be early adopters of new labor-saving technology. In other words, increasing domination of markets by large enterprises could act directly to accelerate both job automation and de-skilling in the service sector. There’s a very real risk that the convergence of all these forces will have a significant dampening effect on the re-generation of the low-wage service jobs that have been a primary engine of American job creation in recent years, and this has the potential to make sustained recovery from the current crisis all the more difficult.

白领自动化浪潮即将来临……为什么教每个人编程并不是解决方案

THE COMING WAVE OF WHITE COLLAR AUTOMATION… AND WHY TEACHING EVERYONE TO CODE IS NOT A SOLUTION

工作自动化的幽灵通常会让人联想到在工厂或仓库里辛苦工作的工业机器人。传统观点认为，虽然低工资、受教育程度较低的蓝领工人面临着来自技术的严重威胁，但受过至少学士学位教育的知识型工人——换句话说，任何以智力而非体力劳动为主的职业——仍然处于相对安全的地位。然而，现实情况是，随着人工智能的发展和越来越广泛的应用，白领工作，尤其是那些以相对常规的信息分析、操纵、提取或交流为中心的工作，将受到威胁。

The specter of job automation typically conjures up images of industrial robots toiling in factories or warehouses. The conventional wisdom suggests that, while lower-wage, less-educated, blue collar workers face a dire threat from technology, knowledge workers educated with at least a bachelor’s degree—or in other words, anyone whose occupation consists of tasks that are primarily of an intellectual, rather than a manual, nature—remain on relatively safe ground. The reality, however, is that white collar jobs, and in particular those that center on relatively routine analysis, manipulation, extraction or communication of information, are going to be squarely in the sights as artificial intelligence advances and is deployed ever more widely.

事实上，在许多情况下，从事信息导向工作的白领专业人士可能比从事需要对环境进行物理操控的职业的受教育程度较低的工人更容易被技术取代。这是因为这些角色的自动化不需要昂贵的机器，在机器视觉或机器人灵活性等领域也不存在需要克服的挑战。相反，消除占用这些工人时间的许多任务只需要足够强大的软件。而消除白领工作的动机进一步被这些高技能工人的工资通常比蓝领工人高得多这一事实所放大。正如我们已经看到的，近一半的应届大学毕业生就业不足，这在某种程度上可能是由于技术对更常规的入门级职位的影响，而这些职位传统上是职业成功的第一步。

Indeed, in many cases, white collar professionals in information-oriented jobs may prove to be even more vulnerable to displacement by technology than less educated workers in occupations that require physical manipulation of the environment. This is because automation of these roles requires no expensive machinery and there are no challenges to overcome in areas like machine vision or robotic dexterity. Instead, eliminating many of the tasks that occupy the time of these workers requires only sufficiently powerful software. And the incentive to eliminate white collar work is further amplified by the fact that these higher-skilled workers are generally paid significantly more than their blue collar counterparts. As we’ve already seen, nearly half of recent college graduates are underemployed, and this is likely driven to some degree by the impact of technology on the more routine entry-level positions that traditionally have offered the first rung on the ladder to professional success.

尽管最危险的工作仍将是那些较为常规的工作，但重要的是要认识到，随着人工智能的不断发展，可以自动化的任务和被视为安全的任务之间的界限肯定是动态的，并将不断变化，涵盖越来越多的活动。以前，知识型活动的自动化需要计算机程序员制定分步程序，明确阐述每个操作和决策。这往往将软件自动化限制在真正常规和重复的任务上，通常是在文书领域，如一般簿记或应付账款和应收账款。然而，机器学习的兴起意味着算法现在可以自由地编写自己的计算机程序，通过翻阅大量数据并找到通常超出人类直接感知的模式和相互关系。换句话说，机器学习的本质就是将曾经被视为本质上非常规的任务转变为现在易于自动化的活动。

Though the jobs most at risk will continue to be those that are more routine, it’s important to realize that the line between tasks that can be automated and those that are perceived as safe is certain to be dynamic and will shift relentlessly to encompass ever more activities as artificial intelligence continues to advance. Previously, the automation of a knowledge-based activity would have required a computer programmer to lay out a step-by-step procedure, with each action and decision explicitly articulated. This tended to limit software automation to truly routine and repetitive undertakings, often in clerical areas like general bookkeeping or accounts payable and receivable. The rise of machine learning, however, means that algorithms are now turned loose to essentially write their own computer programs by churning through reams of data and finding patterns and interrelationships that are often beyond direct human perception. In other words, it is the essence of machine learning to transform tasks that were once perceived as inherently non-routine into activities that are now susceptible to automation.

已经有许多例子表明，软件自动化（通常结合机器学习）开始侵入各种白领职业的活动。例如，在法律领域，智能算法现在审查文件以确定是否需要将其纳入法律发现过程，而人工智能系统在法律研究方面也越来越熟练。预测算法分析历史数据并评估一切的可能性，从最高法院案件的结果到某份合同将来可能被违反的可能性。换句话说，人工智能已经开始影响判断驱动的活动，而这些活动曾经只是最有经验的律师的职权范围。大型媒体组织越来越依赖于通过分析数据流、识别其包含的故事，然后自动生成叙述文本来自动化基本新闻报道的系统。像彭博社这样的公司使用这些系统几乎可以即时创建涵盖公司收益报告的新闻文章。随着人工智能处理自然语言的能力不断提高，几乎任何类型的用于内部或外部组织沟通的常规写作都将越来越容易受到自动化的影响。银行和保险等行业的分析工作可能特别容易受到所有这些影响。例如，富国银行 2019 年的一份报告预测，由于技术进步，未来十年美国银行业将有大约 20 万个工作岗位消失。26^自动化对华尔街的影响已经显而易见，曾经熙熙攘攘、混乱不堪的交易大厅现在基本上充斥着机器的轻柔嗡嗡声。到 2019 年，主要证券交易所只配备了交易大厅某些区域的少数人员。27^冠状病毒大流行表明，随着交易所迅速转向完全电子交易，即使是这少数坚持者也不再必不可少。

There are already numerous examples of how software automation, often incorporating machine learning, is beginning to encroach on activities across a wide variety of white collar occupations. In the field of law, for example, smart algorithms now review documents to determine if they need to be included in the legal discovery process, and artificial intelligence systems are becoming increasingly adept at legal research. Predictive algorithms analyze historical data and assess the probability of everything from the outcome of cases before the Supreme Court to the likelihood that a particular contract might someday be breached. In other words, AI is already beginning to impact judgment-driven activities that would have once been the purview of only the most experienced attorneys. Large media organizations increasingly rely on systems that automate basic journalism by analyzing a stream of data, identifying the story it contains and then automatically generating narrative text. Companies like Bloomberg use these systems to create news articles covering corporate earnings reports almost instantaneously. As the ability of artificial intelligence to handle natural language improves, it’s likely that nearly any type of routine writing intended for either internal or external organizational communication will be increasingly susceptible to automation. Analytical jobs in industries like banking and insurance are likely to be especially vulnerable to all of this. A 2019 report from Wells Fargo, for example, predicted that about 200,000 jobs in the U.S. banking industry would evaporate as the result of advancing technology over the next decade.²⁶ The impact of automation on Wall Street is already evident, and what were once bustling and chaotic trading floors are now largely filled with the soft hum of machines. By 2019, the major stock exchanges were staffed by only small groups of people relegated to certain areas of the trading floor.²⁷ The coronavirus pandemic has demonstrated that even these few holdouts are no longer essential, as the exchanges moved rapidly to fully electronic trading.

提供客户服务或技术支持的呼叫中心是另一个明显容易被颠覆的领域。人工智能自然语言处理能力的快速发展正在产生能够通过语音通信技术和在线聊天机器人自动完成更多此类工作的应用程序。当然，这些工作已经很容易被外包。然而，随着技术的进步，印度和菲律宾等低工资国家的许多呼叫中心工作正被自动化所取代。响应客户服务查询是一项在许多方面非常适合机器学习的任务。客户和呼叫中心工作人员之间的每次互动都会生成一组丰富的数据，包括提出的问题、提供的答案以及互动是否完全解决了问题。机器学习算法可以处理数千次这样的互动，并迅速熟练地响应那些往往会反复出现的大量查询。一旦系统到位，随着越来越多的客户电话打进来，算法就会变得越来越智能。实际上有数十家初创公司提供人工智能聊天机器人来自动化客户服务。其中许多都定位于医疗保健或金融服务等特定领域。28^随着这些技术的不断发展，呼叫中心的人员配备可能会下降，因为最终只有在最具挑战性的客户互动中才需要人工接线员。

Call centers that provide customer service or technical support are another area that is clearly ripe for disruption. Rapid advances in the natural language processing capabilities of artificial intelligence are producing applications that can automate ever more of this work via both voice communication technology and online chatbots. These jobs, of course, were already highly vulnerable to offshoring. As technology has advanced, however, many of the call center jobs in lower-wage countries like India and the Philippines are being vaporized by automation. Responding to customer service queries is a task that in many ways is ideally suited to machine learning. Each interaction between a customer and a call center worker generates a rich set of data, including the question asked, the answer provided and whether the interaction fully resolved the issue. Machine learning algorithms can churn through thousands of these interactions and quickly become proficient at responding to the significant fraction of queries that tend to come up again and again. And once the system is in place, and as more customer calls come in, the algorithms get smarter and smarter. There are literally dozens of startup companies offering AI-powered chatbots to automate customer service. Many of these are positioned in specific sectors such as healthcare or financial services.²⁸ As these technologies continue to advance, call center staffing is likely to fall as things eventually reach a point where a human operator is required only for the most challenging customer interactions.

编写计算机代码的能力通常被视为解决技术就业市场混乱的灵丹妙药。那些在新闻业甚至煤矿开采等行业失业的人被建议“学习编码”。编码学院如雨后春笋般涌现，并且有许多提议要求在高中甚至更早的时候将计算机编程课程作为必修课。但事实上，编写计算机代码肯定会受到破坏其他类型白领工作的同样力量的影响。与呼叫中心一样，外包通常是自动化的前沿，许多常规软件开发已经外包给低工资国家，尤其是印度。几乎所有大型科技公司都在对自动化计算机编程的工具进行大量投资。例如，Facebook 开发了一种名为 Aroma 的工具，它是一种人工智能驱动的计算机编程“自动完成”工具，利用庞大的公共领域计算机代码数据库。29 ^DARPA 还资助了自动化计算机代码开发、调试和测试的研究。即使是 OpenAI 的 GPT-3（一种通过从互联网上提取的大量文档进行训练的通用语言生成系统）也能够完成一些常规的编程任务^。30

The ability to write computer code is often presented as a kind of panacea for technological job market disruption. Those losing jobs in industries like journalism or even coal mining have been advised to “learn to code.” Coding academies have sprung up, and there have been many proposals to make computer programming classes mandatory in high school, or even earlier. In truth, however, writing computer code is certain to be subject to the same forces that will disrupt other types of white collar work. As with call centers, outsourcing is often the leading edge of automation, and much routine software development is already offshored to lower-wage countries, especially India. Nearly all the major tech companies have been making significant investments in tools that automate computer programming. Facebook, for example, has developed a tool called Aroma that works as a kind of AI-powered “autocomplete” for computer programming that leverages a huge database of public domain computer code.²⁹ DARPA has also funded research into automating the development, debugging and testing of computer code. Even OpenAI’s GPT-3, a general language generation system trained on a vast number of documents extracted from the internet, is able to complete some routine programming tasks.³⁰

最重要的是，虽然学习计算机编程肯定是一项有用且有益的工作，但掌握这项技能就能保证获得一份体面工作的日子即将结束。对于其他许多白领职业来说，情况也将如此。随着技术开始侵蚀这些受教育程度更高、收入更高的工人，不平等现象可能会变得更加严重，拥有大量资本的少数精英阶层将其他人拉开距离。随着收入更高的工人受到越来越大的影响，这将进一步削弱消费者支出和强劲经济增长的潜力。然而，一个可能的好处是，收入更高的知识工作者比在工厂或低薪服务行业工作的同行拥有更大的政治权力。因此，对白领工作的影响实际上可能有助于激发对应对技术颠覆就业市场的政策反应的支持。

The bottom line is that though learning to program a computer may certainly be a useful and rewarding undertaking, the days when acquiring that skill guarantees a decent job are coming to a close. And the same will be true for a wide spectrum of other white collar occupations. As technology begins to encroach on even these more educated and highly paid workers, inequality is likely to become ever more top heavy, with a tiny elite that owns vast amounts of capital pulling away from everyone else. As better-paid workers are increasingly impacted, this will further undermine consumer spending and the potential for robust economic growth. One possible upside, however, is that better-paid knowledge workers wield far more political power than their counterparts who work in factories or low-wage service occupations. As a result, the impact on white collar jobs may actually help to galvanize support for a policy response to technological disruption of the job market.

哪些工作最安全？

WHICH JOBS ARE SAFEST?

过去几年，我几乎走遍了各大洲，就人工智能和机器人技术对就业市场的潜在影响做了几十次演讲。无论我身处哪个国家，我发现听众最常问的问题几乎都是一样的：哪些工作最有可能安全，我应该建议孩子学习哪些专业？一般的答案可能有点明显，也不太令人满意：避免从事那些基本上是例行公事、本质上可预测的工作。这些领域显然将在短期内受到人工智能自动化的最大影响。另一种表达方式可能是“避免从事无聊的工作”。如果你每天都在工作，面对新的挑战，并且在工作中不断学习，那么你很可能处于领先地位，至少在可预见的未来是这样。另一方面，如果你花费大量时间一遍又一遍地编写相同类型的报告、演示文稿或分析，你可能应该开始担心，并开始考虑调整你的职业发展轨迹。

Over the past few years, I’ve traveled to nearly every continent and given dozens of presentations on the potential impact of artificial intelligence and robotics on job markets. Regardless of the country I happen to be in, I’ve found that the most common questions I receive from the audience are nearly always the same: what jobs are most likely to be safe, and what fields should I advise my children to study? The general answer is perhaps a bit obvious and unsatisfying: avoid jobs that tend to be fundamentally routine and predictable in nature. These are clearly the areas that will see the most significant near-term impact of AI-powered automation. Another way to phrase this might be “avoid jobs that are boring.” If you’re coming to work and facing new challenges every day and if you’re constantly learning on the job, then you’re probably well positioned to stay ahead of technology, at least for the foreseeable future. If, on the other hand, you spend a lot of time cranking out the same kinds of reports, presentations or analyses again and again, you should probably start to worry—and begin to think about adjusting your career trajectory.

更具体地说，我认为在短期和中期内最不容易被自动化取代的工作分为三个大类。首先，真正具有创造性的工作可能相对安全。如果你能跳出固有思维模式，想出创新策略来解决无法预见的问题，或者创造一些真正新颖的东西，那么我认为你将能够很好地利用人工智能作为工具。换句话说，这项技术更有可能成为你的补充，而不是取代你。可以肯定的是，关于建造创造性机器的大量研究正在进行中，人工智能也将不可避免地开始侵占创造性工作。智能算法已经可以绘制原创艺术作品、提出科学假设、创作古典音乐和生成创新的电子设计。DeepMind 的 AlphaGo 和 AlphaZero 为专业围棋和国际象棋比赛注入了新的活力和创造力，因为这些系统代表了真正的外星智能，经常采用令人类专家惊讶的非常规策略。然而，我认为在可预见的未来，人工智能将被用来放大而不是取代人类的创造力。

More specifically, I think the jobs least susceptible to automation in the near to intermediate term fall into three general areas. First, jobs that are genuinely creative in nature are likely to be relatively safe. If you’re thinking outside the box, coming up with innovative strategies to solve unforeseen problems or building something genuinely new, then I think you will be well positioned to leverage artificial intelligence as a tool. In other words, the technology is much more likely to compliment you than it is to replace you. To be sure, significant research into building creative machines is underway, and AI will inevitably begin to encroach on creative work as well. Already smart algorithms can paint original works of art, formulate scientific hypotheses, compose classical music and generate innovative electronic designs. DeepMind’s AlphaGo and AlphaZero have injected new energy and creativity into professional Go and chess competitions because the systems represent truly alien intelligences, often adopting unconventional strategies that astonish human experts. However, I think that for the foreseeable future artificial intelligence will be used to amplify, rather than replace, human creativity.

第二个安全领域包括那些重视与他人建立有意义且复杂的关系的工作。例如，这包括护士与患者之间可能存在的那种富有同情心和关爱的关系，或者提供复杂建议的商人或顾问与客户之间可能存在的那种关系。需要注意的是，我指的并不是那些需要微笑和友好对待客户的短期服务接触，而是那些需要更深入、更复杂的人际互动的服务接触。人工智能也正在侵入这个领域；正如我们在第 3 章中看到的那样，聊天机器人已经可以提供基本的心理健康治疗，人工智能在感知、响应和模拟人类情感方面的能力将继续取得重大进展。但我认为，机器还需要很长时间才能与人建立真正复杂的多维关系。

A second safe area consists of those jobs that put a premium on building meaningful and complex relationships with other people. This would include, for example, the kind of empathetic, caring relationship that a nurse might have with a patient, or that a business person or consultant offering sophisticated advice might develop with a client. It’s important to note that I’m not referring so much to short-term service encounters that involve smiling and being friendly with customers, but rather those that require deeper and more complex interpersonal interactions. Once again, AI is also encroaching into this area; as we saw in Chapter 3, chatbots can, for example, already provide rudimentary mental health therapy, and there will continue to be significant advances in AI’s ability to perceive, respond to and simulate human emotion. I think it will be a long time, however, before machines become capable of developing truly sophisticated, multidimensional relationships with people.

第三类安全工作包括在不可预测的环境中需要高度机动性、灵活性和解决问题能力的职业。护士和老年护理助理也属于这一类，水管工、电工和机械师等技术性职业也属于这一类。制造能够自动完成这类工作的廉价机器人可能还远在未来。这些技术性职业通常代表着那些选择不接受大学教育的人的一些最佳机会。在美国，我认为我们应该更加重视可获得的职业培训或学徒制，为年轻人做好这些机会的准备，而不是仅仅推动越来越多的高中毕业生上大学。

The third general category of safe jobs includes occupations that require significant mobility, dexterity and problem-solving skills in unpredictable environments. Nurses and elder care assistants would also fall into this category, as would skilled trade occupations like plumbers, electricians and mechanics. Building affordable robots capable of automating work of this type likely lies far in the future. These skilled trade jobs will generally represent some of the best opportunities for those who choose not to pursue a college education. In the United States, I think we should be putting far more emphasis on accessible vocational training or apprenticeships that prepare young people for these opportunities rather than simply pushing ever more high school graduates to attend college or university.

然而，最重要的因素可能不是你选择哪种职业，而是你如何定位自己。随着人工智能的发展，在广阔的就业市场中，主要由常规“基本”活动组成的工作可能会消失，而那些专注于需要创造性技能的领域或能够利用广泛的专业网络为组织增值的人将脱颖而出。换句话说，在以前机会更为同质的职业中，可能会出现你在运动员或演艺人员身上看到的那种赢家通吃或超级明星效应。即使人工智能不断发展，拥有强大法庭技能或为公司带来业务的客户关系的律师也可能继续表现出色。另一方面，主要从事法律研究或合同分析的律师可能处于不太光明的境地。

The most important factor, however, may not be so much which occupation you choose but how you position yourself within it. As artificial intelligence advances, it’s likely that across broad swathes of the job market, jobs consisting largely of routine “nuts and bolts” activities will evaporate, while those who focus in areas requiring creative skills or who can leverage extensive professional networks in ways that add value to organizations will rise to the top. In other words, there’s likely to be something of a winner-take-all or superstar effect of the kind you see among athletes or entertainers imposed on occupations that were previously more homogeneous in terms of opportunity. A lawyer with strong courtroom skills or the client connections that bring business to the firm will likely continue to do well even as artificial intelligence advances. An attorney who, on the other hand, toils away mostly on legal research or contract analysis may be in a less promising situation.

作为个人，适应这种情况的最佳方式可能是选择一份你真正喜欢的职业——你热爱的职业——因为这会增加你在该领域脱颖而出并成为异军突起的机会。展望未来，仅仅因为该领域传统上提供了很多工作机会而选择一份职业可能不是一个好选择。当然，问题在于，这对某个人来说可能是个好建议，但它不是一个系统性的解决方案。随着这些转变的展开，很多人很可能会落后，最终，我认为我们需要制定政策来解决这一现实。

The best way for you as an individual to adapt to this situation is probably to select an occupation that you genuinely enjoy—something that you’re passionate about—because this will increase your chances of excelling and becoming an outlier in the field. Going forward, choosing an occupation simply because the field has traditionally provided lots of jobs may not be such a good bet. The problem, of course, is that this may be good advice for a given individual, but it is not a systemic solution. Many people will, in all likelihood, be left behind as these transitions unfold, and ultimately, I think we will need policies to address that reality.

经济上行

THE ECONOMIC UPSIDE

尽管人工智能对就业市场和经济不平等的潜在影响确实令人担忧，但毫无疑问的是，这项技术将为经济和社会带来巨大利益。自动化程度的提高将提高生产效率，并直接导致商品和服务价格下降。换句话说，人工智能将成为缓解并最终消除贫困的重要工具，因为它使人们赖以生存所需的一切变得更加丰富和负担得起。人工智能在研究、设计和开发中的应用将带来全新的产品和服务，而这些产品和服务原本是不可想象的。新药和新疗法将带来巨大的经济效益，同时提高几乎所有人的福祉。

Though the potential impact of artificial intelligence on the job market and on economic inequality are real concerns, there is also no doubt that the technology is poised to deliver enormous benefits across both the economy and society. Increased automation will boost the efficiency of production and lead directly to lower prices for goods and services. In other words, AI will be a critical tool for alleviating—and eventually eliminating—poverty by making all the things that people need to thrive more abundant and affordable. Artificial intelligence deployed in research, design and development will lead to entirely new products and services that otherwise might have been unimaginable. New medicines and treatments will lead to dramatic economic benefits while increasing the well-being of nearly everyone.

2018 年末发布的两份报告，一份来自麦肯锡全球研究院³¹ ，另一份来自咨询公司普华永道^32，均强烈指出人工智能将在 2030 年之前极大地促进全球经济。麦肯锡的分析预测人工智能将为全球经济产出增加约 13 万亿美元，而普华永道的估计为 15.7 万亿美元。换句话说，人工智能很可能在未来十年左右增加新的全球经济价值，大致相当于中国目前 14 万亿美元的 GDP。麦肯锡的分析表明，这些收益将以 S 曲线的方式实现——“由于学习和部署 [人工智能] 涉及大量成本和投资，起步缓慢，但随后在竞争的累积效应和互补能力的提高的推动下加速。” ³³ 到 2030 年，我们很可能会发现自己正处于曲线的陡峭、加速部分，技术和与之相关的经济收益都在快速发展。

Two reports released in late 2018, one from the McKinsey Global Institute³¹ and the other from the consulting firm PwC,³² both make a strong argument that artificial intelligence will deliver a massive boost to the global economy by the year 2030. McKinsey’s analysis projects that AI will add about $13 trillion to economic output worldwide, while PwC’s estimate comes in at $15.7 trillion. In other words, AI is likely poised to add new global economic value roughly equivalent to China’s current $14 trillion GDP over the next decade or so. McKinsey’s analysis suggests that these gains will arrive in a fashion that traces an S-curve—“a slow start due to substantial costs and investment associated with learning and deploying [artificial intelligence], but then an acceleration driven by the cumulative effect of competition and an improvement in complementary capabilities.”³³ By 2030, it’s likely we’ll find ourselves on the steep, accelerating portion of the curve with both the technology and the economic gains associated with it progressing rapidly.

这些估计在很大程度上未能反映出人工智能长期最显著的好处。正如我在第 3 章中指出的那样，人工智能最重要的前景是它可以帮助我们摆脱技术停滞的时代。如果人工智能能让我们在广泛的科学、工程和医学领域启动创新，那么我们投资的潜在回报将是惊人的。也许最重要的是，迫切需要扩大我们的集体智慧和创造力，以便我们能够应对必然会面临的艰巨挑战——包括从气候变化到新的清洁能源，再到应对下一次流行病等一切挑战。这些因素很难用经济分析来量化，但我认为，仅凭这些因素，人工智能就成为一种不可或缺的工具，我们根本不能放弃它，即使它伴随着前所未有的经济和社会风险。

These estimates largely fail to capture the most dramatic benefits from artificial intelligence over the long term. As I argued in Chapter 3, the single most important promise of AI is that it can help us escape from our age of technological stagnation. If artificial intelligence allows us to jump-start innovation across a broad range of scientific, engineering and medical fields, the potential return on our investment will be staggering. What is perhaps most important is the critical need to amplify our collective intelligence and creativity in ways that will allow us to address the daunting challenges that are sure to confront us—including everything from climate change to new sources of clean energy to managing the next pandemic. These things are difficult to quantify with economic analysis, but I would argue they alone make artificial intelligence an indispensable tool that we simply cannot afford to leave on the table, even as it comes coupled with unprecedented economic and social risks.

我们面临的关键挑战是找到解决技术性失业和不平等加剧等弊端的方法，同时继续投资人工智能并充分利用该技术带来的优势。我们将面临的根本经济挑战是分配问题。人工智能带来的潜在经济收益是不可否认的，但绝对不能保证这些收益会在整个人口中得到广泛或公平的分享。事实上，如果我们不采取任何行动，几乎可以肯定的是，收益将主要集中在收入分配顶端的一小部分人身上，而大部分人口将被抛在后面，甚至可能变得更糟。正如我们所看到的，这反过来可能会削弱广泛的消费需求，抑制生产率增长和经济增长。换句话说，如果不能解决人工智能的经济弊端，很可能会限制该技术优势的充分实现。我认为，要避免这种默认结果，需要采取重大和非常规的政策举措。几十年来一直采用的传统解决方案——职业再培训计划或鼓励更多人上大学——可能已不够用，尤其是考虑到人工智能已经对高技能工作产生了重大影响，而且随着技术变得越来越强大，这种趋势只会越来越强劲。

The key challenge before us is to find ways to address downsides like technological unemployment and increased inequality while continuing to invest in AI and fully leverage the advantages the technology will bring. The fundamental economic challenge we will face is one of distribution. The potential economic gains associated with artificial intelligence are undeniable, but there is absolutely no guarantee that these benefits will be shared broadly or fairly across the population. Indeed, if we take no action at all, it seems a near certainty that the gains will accrue overwhelmingly to a small sliver of people at the top of the income distribution, while the bulk of the population will be largely left behind or potentially even made worse off. And as we’ve seen, this in turn could erode broad-based consumer demand and dampen down both productivity gains and economic growth. In other words, a failure to address the economic downsides of AI may well limit the full realization of the technology’s upside. Avoiding this default outcome will, I think, require dramatic and unconventional policy initiatives. Traditional solutions that have been deployed for decades—job retraining programs or pushing ever more people to attend college—are unlikely to be sufficient, especially given that artificial intelligence is already having a significant impact on higher-skill jobs, and this trend will only gain traction as the technology becomes ever more capable.

解决分配问题

FIXING THE DISTRIBUTIONAL PROBLEM

在我看来，解决人工智能进步带来的分配挑战最直接、最有效的方法就是给人们钱。换句话说，用某种形式的最低收入保障、负所得税或基本收入来补充所有或大部分人口的收入。最近最受关注的想法是无条件的全民基本收入，即 UBI。作为对人工智能驱动的自动化的政策回应，UBI 的知名度在 2019 年因安德鲁·杨 (Andrew Yang) 的总统候选人资格而大大提高。寻求民主党总统候选人提名的杨主要以每月 1,000 美元的“自由红利”为竞选纲领，该红利将支付给所有美国公民。他的竞选活动获得了显著的关注，这主要是由于网上的大量追随者，他参加民主党辩论将 UBI 推向主流，并首次让大量美国人接触到这个想法。

In my view, the most straightforward and effective way to address the distributional challenge brought about by artificial intelligence advances is simply to give people money. In other words, supplement the incomes of all or the bulk of the population with some version of a guaranteed minimum income, negative income tax or basic income. The idea that has recently gained the most traction is an unconditional universal basic income, or UBI. The visibility of UBI as a policy response to AI-driven automation was dramatically accelerated in 2019 by the presidential candidacy of Andrew Yang. Yang, who sought the Democratic nomination for president, ran primarily on a platform of a $1,000 per month “Freedom Dividend” that would be paid to all U.S. citizens. His campaign gained remarkable traction largely as the result of a vibrant online following, and his participation in the Democratic debates pushed UBI into the mainstream and exposed large numbers of Americans to the idea for the first time.

无条件基本收入的主要优势之一是，由于它支付给每个人，无论就业状况如何，它不会破坏受助人工作或从事产生额外收入的创业活动的积极性。换句话说，它避免了传统安全网计划的最大问题之一：容易造成贫困陷阱。由于一旦受助人找到工作并开始赚取收入，失业保险或福利金等计划就会逐步取消或完全取消，因此人们寻找工作的积极性会大大降低。即使是接受一份低薪工作也会立即危及一个人现有的收入。因此，人们经常陷入对安全网计划的依赖，几乎看不到采取小步骤迈向更美好未来的具体激励。相比之下，全民基本收入不受就业影响，因此任何选择工作或创办一家产生额外收入的小企业的人，总是比那些只是坐在家里领取每月 UBI 付款的人过得更好。 UBI 创造了绝对的收入底线，但人们总是有强烈的动机去赚更多的钱。尽管有这个优势，但对于许多人来说，他们对于直接把钱交给人们，或者用一些人的话来说，“付钱让人们活下去”，有着强烈的心理厌恶，这种态度很可能会继续对 UBI 的实际实施造成重大的政治障碍。

One of the primary advantages of an unconditional basic income is that, because it is paid to everyone regardless of employment status, it doesn’t destroy the incentive for recipients to work or engage in entrepreneurial activity that generates additional income. In other words, it avoids one of the biggest problems with traditional safety net programs: the tendency to create a poverty trap. Because programs like unemployment insurance or welfare payments are phased out or eliminated entirely once the recipient finds a job and begins to earn income, there can be a powerful disincentive to seek employment. Accepting even a low-paying job puts the person’s existing income at immediate risk. As a result, people often get trapped into dependency on safety net programs and see little in the way of a concrete incentive to take small steps toward a better future. A universal basic income, in contrast, is unaffected by employment, and therefore anyone who chooses to work or perhaps start a small business that generates extra income will always be better off than the person who simply sits at home and collects the monthly UBI payment. The UBI creates an absolute income floor, but there always remains a strong incentive to earn more. Despite this advantage, there is, for many people, a strong psychological aversion to the idea of simply handing money to people, or as some have put it, “paying people to be alive,” and this attitude is likely to continue to create a significant political impediment to actual implementation of a UBI.

当然，还有其他政策选择，其中最常提到的就是就业保障。让政府成为任何需要工作的人的最后雇主的想法表面上看起来很有吸引力，但我认为它有重大的缺点。就业保障的普及程度远不及基本收入，而且不可避免地会有很多最需要帮助的人被排除在外。这样的制度需要庞大、昂贵且可能不断扩大的官僚机构。管理人员需要确保工人确实出现在他们被分配的任何工作中，而且毫无疑问会出现一系列纪律问题，从旷工到表现不佳到“我也是”的情况。任何旨在惩戒或解雇未能达到规定标准的工人的政策都将充满争议，很可能会被指控歧视或不平等待遇。最终，政府要么必须解雇那些表现不佳或违反规定的人（这将使受影响的个人被排除在安全网之外），要么就业计划实际上会变成非常昂贵且效率低下的基本收入计划。创造的职位中很大一部分很可能是“垃圾工作”，与基本收入计划不同，就业保障会直接吸引工人离开私营部门的生产效率更高的职位。相比之下，基本收入几乎不需要官僚机构，并将利用政府现有的通过社会保障等计划发放支票的能力。

To be sure, there are other policy alternatives, one of the most commonly cited being a jobs guarantee. The idea of having the government become the employer of last resort for anyone who needs a job may seem superficially attractive, but I think it has significant disadvantages. A jobs guarantee would be far less universal than a basic income, and inevitably many of the people who are most in need of assistance would be left out. Such a system would require a massive, costly and likely ever-expanding bureaucracy. Managers would need to ensure that workers actually show up to do whatever job they’ve been assigned, and there would doubtless be a whole slew of disciplinary issues ranging from absenteeism to poor performance to “me too” situations. Any policies geared toward disciplining or terminating workers who failed to meet the specified standards would be fraught with controversy and quite possibly with accusations of discrimination or unequal treatment. Ultimately, the government would either have to fire those who underperformed or broke the rules—which would exclude any impacted individuals from the safety net—or the jobs program would effectively become the equivalent of a very expensive and inefficient basic income scheme. A large fraction of the positions created would in all likelihood be “bullshit jobs,” and unlike a basic income program, a jobs guarantee would directly attract workers away from more productive positions in the private sector. In contrast, a basic income requires little in the way of a bureaucracy and would take advantage of the government’s existing competence at sending out checks via programs like Social Security.

虽然我认为基本收入最终是解决人工智能普及后出现的分配问题的最佳整体解决方案，但它绝不是万能药。相反，我认为 UBI 是建立更有效、政治上可接受的解决方案的基础。最重要的问题是，虽然基本收入将钱送到人们手中，但它本身并不能复制传统工作的其他重要品质。有意义的工作提供了一种目标感和尊严感。它占用时间，并激励人们努力工作，力求出类拔萃，以期获得加薪或晋升。获得一份好工作的愿望也是个人继续接受教育和培训的至关重要的动力。

While I think a basic income is ultimately the best overall solution to the distributional issues that will emerge as artificial intelligence becomes ubiquitous, it is by no means a panacea. Rather, I view UBI as a foundation upon which to build a more effective and politically palatable solution. The most important problem is that though a basic income delivers money into the hands of people, it alone does not replicate the other important qualities associated with a traditional job. A meaningful job provides a sense of purpose and dignity. It occupies time and creates an incentive to work hard and to excel in the hope of obtaining a raise or a promotion. The desire to obtain a good job is also a critically important incentive for individuals to pursue further education and training.

我认为可以修改基本收入计划，使其至少部分复制这些特点。自从 2009 年出版我的第一本书《 隧道中的灯光：自动化、加速技术和未来经济》以来，我一直倡导直接纳入激励措施的基本收入计划。虽然每个人都应该获得一些最低限度的保证金，但我认为也应该有机会通过从事某些活动来多赚一些钱。迄今为止最重要的激励措施应该是继续深造。想象一个世界，每个人从 18 岁或 21 岁开始每月都收到完全相同的 UBI 付款。在这种情况下，有辍学风险的高中生可能没有什么理由努力获得文凭。毕竟，不管怎样，每月的薪水都是一样的。如果——似乎已经是这样——获得文凭不足以找到一份好工作，那为什么还要继续上学呢？我认为这种抑制措施将是灾难性的，会引发人们对受教育程度较低的人口的担忧，尽管我们面临的未来正变得越来越复杂，充满了艰难的挑战和权衡。因此，为什么不简单地给高中毕业生多付一点钱呢？将激励措施纳入基本收入计划的想法可能会扩展到包括更先进的教育，甚至包括社区服务工作等其他领域。最终的愿景是创造机会，让人们有意义地度过自己的时间并获得成就感。也许最重要的是，那些受到激励继续深造的人将增加他们通过就业或创业活动获得更多机会的可能性。随着人工智能的应用越来越广泛，它很可能会提供强大的工具，个人可以利用这些工具创办小企业或通过自由职业机会赚取收入，但利用这些机会至少需要达到最低的教育门槛。保持强大的激励，让社会各个阶层的每个人都在力所能及的范围内努力获得最高的教育水平，这应该是我们最重要的目标之一。

I believe it’s possible to modify a basic income program so it, at least in part, replicates some of these qualities. Since the publication of my first book, The Lights in the Tunnel: Automation, Accelerating Technology and the Economy of the Future, in 2009, I’ve advocated for a basic income scheme that directly incorporates incentives. Though everyone should receive some minimal guaranteed payment, I think there should also be opportunities to earn somewhat more by pursuing certain activities. The most important incentive by far should be to pursue further education. Imagine a world where everyone receives exactly the same monthly UBI payment beginning at age eighteen or perhaps twenty-one. In this situation, a high school student at risk of dropping out of school might see very little reason to work hard to obtain that diploma. The monthly check, after all, is going to be the same no matter what. And if—as it seems to already be the case—getting that diploma will not be sufficient to obtain a good job anyway, then why stay in school? I think this kind of disincentive would be disastrous and would raise the specter of a less-educated population even as we face a future that is becoming vastly more complex and fraught with difficult challenges and trade-offs. Therefore, why not simply pay a bit more to those who graduate from high school? This idea of building incentives into a basic income program might be expanded to include more advanced education and perhaps other areas such as community service work. The ultimate vision is to create opportunities that would offer people meaningful ways to spend their time and achieve a sense of accomplishment. Perhaps most importantly, those who act on the incentive to pursue further education will increase the odds that they will be able to access still more opportunities through employment or entrepreneurial activity. It’s likely that as artificial intelligence is deployed ever more widely, it will provide powerful tools that individuals will be able to leverage to start a small business or perhaps generate income through freelancing opportunities, but taking advantage of these opportunities will require achieving at least a minimum educational threshold. Preserving a strong incentive for everyone at every level of our society to strive for the highest level of education within their ability should be one of our most important goals.

UBI 的另一个主要问题是它非常昂贵。无条件地向每个美国成年人分配收入将花费数万亿美元，选民可能会对每月向已经富裕的人发放支票的想法感到反感。我认为可能有机会在收入较高时有效地逐步取消 UBI，而不会影响工作的积极性。做到这一点的最佳方法可能是仅针对“被动收入”对 UBI 进行经济状况调查。如果您已经拥有大量自动收入，无需您工作或采取行动（如果您收到养老金、社会保障或大量投资收入），那么我认为相应地逐步取消或取消 UBI 付款是合理的。工作或直接管理企业产生的主动收入不会影响 UBI，除非收入水平非常高。许多人会认为这不公平，但基本收入背后的想法毕竟是为每个人提供至少最低限度的保证收入底线。如果您已经可以获得这种付款，那么可以说您不需要 UBI。没有任何政策举措能够让世界完全公平。我们现实中能期望的最好结果就是制定一个计划，以缓和不平等现象，消除最严重的物质匮乏，并确保消费者有足够的收入来继续推动经济增长。

Another major problem with UBI is that it’s very expensive. Distributing income unconditionally to every adult American will cost trillions, and voters are likely to recoil at the idea of sending monthly checks to the already well-off. I think there may be opportunities to effectively phase out the UBI at higher incomes, without impacting the incentive to work. The best way to do this might be to means-test a UBI against only “passive income.” If you already have significant income that comes to you automatically, without any need for work or action on your part—if you receive a pension, Social Security or substantial investment income—then I think it would be reasonable to phase out or eliminate a UBI payment accordingly. Active income that results from work or direct management of a business would not affect the UBI, except perhaps at a very high income level. Many would perceive this as unfair, but the idea behind a basic income is, after all, to provide everyone with at least a minimal guaranteed income floor. If you already have access to such a payment, then arguably you don’t need the UBI. No policy initiative is ever going to make the world completely fair. The best we can realistically hope for is a program that moderates inequality, eliminates the most dire forms of material deprivation and ensures that consumers have the income they need to continue driving economic growth.

当然，所有这些想法都面临着自身的挑战。如果我们将激励措施纳入基本收入计划，那么究竟谁来决定这些激励措施是什么？对许多人来说，这将立即引发一个专横的保姆国家破坏选择自由并将其融入我们日常生活的担忧。不过，我认为应该能够就至少一套明显对个人和整个社会都有利的最低限度的激励措施达成某种广泛的共识，而且我再次认为追求教育显然是最重要的。一个相关的担忧是基本收入计划的政治化。很容易想象未来几乎每个政客都在“我将增加你的每月 UBI 支付”的平台上竞选。出于这个原因，我认为将基本收入计划的管理从政治进程中移除并将其交给一个按照明确指导方针运作的专门技术机构——换句话说，一个类似于美联储的机构——是非常有意义的。

All these ideas, of course, face challenges of their own. If we incorporate incentives into a basic income scheme, then who exactly gets to decide what those incentives are? For many people, this will immediately raise the specter of an overbearing nanny state undermining freedom of choice and injecting itself into the daily fabric of our lives. Still, I think it ought to be possible to come to some broad agreement on at least a minimal set of incentives that are clearly advantageous both for individuals and for society as a whole, and once again, I think pursuit of education stands apart as clearly being the most important. A related concern would be the politicization of a basic income program. It’s quite easy to imagine a future where nearly every politician runs on a platform of “I will increase your monthly UBI payment.” For this reason I think it would make a lot of sense to remove the administration of a basic income program from the political process and place it in the hands of a dedicated technocratic agency operating according to clear guidelines—in other words, an institution similar to the Federal Reserve.

所有这些并不是说我们应该放弃解决失业、就业不足或不平等加剧等问题的更传统的方法。在未来几年和几十年里，随着人工智能和机器人技术的影响不断加速，我们应该尽一切可能确保最大数量的工人能够成功转型。特别是，我们应该投资社区大学和负担得起的职业培训或学徒计划，这些计划为掠夺性的营利性学校提供了替代方案，而在美国，这些学校目前占据了大部分空间。不过，我认为，最终这种颠覆将是如此之大，以至于这种类型的计划最终将无法满足要求，我们需要采取更多非常规的解决方案。

None of this is to suggest that we should abandon more conventional solutions to unemployment, underemployment or rising inequality. We should do everything possible to ensure that the maximum number of workers are able to successfully transition as the impact from artificial intelligence and robotics accelerates in the coming years and decades. In particular, we should invest in community colleges and affordable vocational training or apprenticeship programs that offer an alternative to the predatory for-profit schools that, in the United States, currently occupy much of this space. Still, I think that eventually the disruption will be of such magnitude that programs of this type will ultimately fall short, and we will need to adopt more unconventional solutions.

基本收入的实现仍将面临巨大的政治障碍，我认为，从现实角度来看，这样的计划可能需要从最低水平开始，然后随着时间的推移逐渐扩大。在实施国家计划之前，我们需要更多数据和更多有关 UBI 的实际经验；因此，我们应该启动旨在找到最佳政策参数的实验。我希望其中一些实验最终能包括我提出的纳入激励措施的想法。基本收入实验产生的数据将使我们能够制定一个能够有效扩展的计划，并有助于确保在日益受人工智能影响的未来实现广泛的繁荣。

The political hurdles standing in the way of a basic income will remain daunting, and I think that realistically such a program may need to begin at a minimal level and then gradually be ramped up over time. Before a national program can be enacted, we need more data and more actual experience with UBI; therefore, we should initiate experiments designed to find the optimal policy parameters. I hope that some of these experiments might eventually include my idea of incorporating incentives. The data generated through basic income experiments will allow us to craft a program that will scale effectively and help ensure broad-based prosperity in a future shaped increasingly by AI.

科技失业和不平等加剧的可能性只是人工智能崛起所带来的主要担忧之一。接下来的两章将重点讨论一系列已经显现或可能随着技术进步而出现的其它危险。

THE POTENTIAL FOR technological unemployment and increasing inequality is just one of the major concerns that comes coupled with the rise of artificial intelligence. The next two chapters will focus on a range of other dangers that are already becoming apparent or are likely to arise as the technology progresses.

脚注

Footnote

ⁱ我们的时间旅行者基于前财政部长兼国家经济委员会主任劳伦斯·萨默斯 (Lawrence Summers) 的观点，他于 2016 年 11 月估计，到 2050 年，四分之一到三分之一的工作年龄男性将退出劳动力市场。（见第 6 章尾注 1。）

ⁱ Our time traveler is based on former Treasury Secretary and Director of the National Economic Council Lawrence Summers, who made the estimate of a quarter to a third of working age men being out of the workforce by 2050 in November 2016. (See endnote 1, Chapter 6.)

³³

第七章

CHAPTER 7 中国与人工智能监控国家的崛起

CHINA AND THE RISE OF THE AI SURVEILLANCE STATE

新疆自治区是中国的西北边疆。该地区面积广阔，大约是德克萨斯州的两倍半，除中国外，还与七个国家接壤：东北部与蒙古接壤，北部与俄罗斯接壤，西部与哈萨克斯坦、吉尔吉斯斯坦、塔吉克斯坦、阿富汗、巴基斯坦和印度接壤。这里的气候和地形十分险峻，主要是崎岖的山脉和沙漠，中间点缀着绿洲城市，该省 2400 万人口中的大多数都聚集在这里。传奇的丝绸之路（或真正的“道路”，因为它实际上是一个路线网络）横穿新疆，使该地区成为东西方贸易的中心，促进了整个欧亚大陆文明的崛起。马可波罗在 13 世纪末曾走过这条路线，他曾见过熙熙攘攘的集市和满载货物的骆驼，与今天在新疆看到的集市和骆驼没有什么不同。

THE XINJIANG AUTONOMOUS REGION CONSTITUTES CHINA’S northwestern frontier. The region is massive—roughly two and half times the size of Texas—and borders seven nations aside from China: Mongolia to the northeast, Russia to the north, and Kazakhstan, Kyrgyzstan, Tajikistan, Afghanistan, Pakistan, and India to the west. The climate and the terrain are formidable—primarily rugged mountains and desert punctuated by the oasis cities where most of the province’s twenty-four million people are clustered. The legendary Silk Road—or really “roads,” as it was actually a network of routes—cut across Xinjiang and made the region central to the East-West trade that contributed to the rise of civilizations throughout Eurasia. Marco Polo traveled this route in the late thirteenth century and would have encountered bustling bazaars and laden camels not unlike those that can be seen in Xinjiang today.

新疆之所以备受关注，并不是因为它悠久的历史，而是因为该地区最大的民族维吾尔族被强加了奥威尔式的未来。在喀什这样的城市，中国建立了一个压迫性的监控国家，由大量警察、实体检查站和先进技术相结合。几乎城市中的每个人都受到持续监控：街道两旁、建筑物上和电线杆上都安装了数千个摄像头。居民在城市中穿行时，会在检查站被拦下，只有通过面部识别系统识别后才允许前进^。1

Xinjiang has been thrust into the limelight not for its rich history, but rather because an Orwellian future has been imposed on the region’s largest ethnic group, the Uyghurs. In cities like Kashgar, China has built an oppressive surveillance state, powered by a combination of massive police presence, physical checkpoints and advanced technology. Virtually everyone in the city is continuously watched: thousands of cameras line the streets, mounted on buildings and clustered on telephone poles. As residents move through the city, they are stopped at checkpoints and allowed to advance only after being identified by facial recognition systems.¹

虽然新疆是中国监控计划的重灾区，但该地区也是技术试验场，这些技术正逐渐在全国范围内推广。预计到 2020 年，中国将安装近 3 亿个摄像头，其中许多摄像头与面部识别技术相连，或采用其他人工智能驱动的跟踪技术，例如根据步态或衣着识别行人。

While Xinjiang is ground zero for China’s surveillance initiatives, the region also acts as a proving ground for techniques and technologies that are gradually being deployed across the entire country. China is expected to have installed nearly 300 million cameras by 2020, many of which are linked to facial recognition technology or feature other AI-driven tracking techniques, such as identifying pedestrians based on their gait or clothing.

在新疆，如果维吾尔人偏离规定的行为或参与禁忌活动（如阅读《古兰经》），就有可能被送往中国在该地区建立的大型“再教育营”。即使在中国其他地区，中国政府也有一个可怕的系统性行为改变愿景，通过部署全面的社会评级系统来实施。最终，一个人生活的几乎所有方面——消费购买、身体活动、社交媒体互动和与他人的交往——都将被监视、记录和分析。这些信息将用于生成每个人的总体社会评级。那些在这个指标上得分低的人将受到惩罚，例如被禁止乘坐公共交通工具或禁止其子女入学。

In Xinjiang, Uyghurs who deviate from prescribed behaviors or engage with forbidden ideas, such as reading the Koran, risk being sent to one of the massive “re-education camps” that China has built in the region. Even in other areas of the country, the Chinese government has a terrifying vision for systemic behavior modification, implemented through the deployment of a comprehensive social rating system. Eventually, nearly all aspects of a person’s life—consumer purchases, physical movements, social media interactions and associations with others—will be surveilled, recorded and analyzed. This information will then be used to generate an overall social rating for each individual. Those who score low on this metric will suffer penalties such as being barred from public transportation or having their children prevented from enrolling in schools.

这一切都因中国迅速崛起成为人工智能研究和开发领域的世界领先者而加速。从某些指标来看，例如从事该领域的计算机科学家和工程师的数量以及发表的研究论文数量，中国已经领先于美国。中国正在大规模投资，并将人工智能作为国家战略要务。中国的领导人似乎既积极参与又知识渊博。2018 年初，中国国家主席习近平在办公室发表电视讲话，背景中可以看到有关人工智能和机器学习的书籍。2^政府还帮助资助了数百家初创公司，其中许多公司的估值达数十亿美元，是当之无愧的技术领导者。

All this is being accelerated by China’s rapid rise to become a world leader in artificial intelligence research and development. By some measures, such as the sheer number of computer scientists and engineers working in the field and the volume of research papers published, China has already taken a lead over the United States. The country is investing massively and has made artificial intelligence a strategic national imperative. Its leaders appear to be both engaged and knowledgeable. In early 2018, Chinese president Xi Jinping gave a televised address from his office, and books on AI and machine learning were spotted in the background.² The government is also helping to fund hundreds of startup companies, many of which are valued at billions of dollars and are clear technology leaders.

随着中国成为全球两大人工智能研发中心之一，中国与美国和西方在这一领域的持续竞争可能会变得更加激烈。中国新兴人工智能产业的很大一部分专注于开发人脸识别和其他监控技术，这些公司不仅在中国，而且在世界各国都找到了热切的客户。而且，正如我们将看到的，基于人工智能的监控技术绝不限于独裁政权。尤其是人脸识别，它正在美国和其他民主国家得到广泛应用，并已引发激烈的争论以及偏见和滥用的指责。随着这项技术变得越来越强大，除非受到严格监管，否则将无处不在，这些问题将变得更加棘手。

As China assumes its role as one of the world’s two primary centers of artificial intelligence research and development, the ongoing competition in this arena with the United States and the West is likely to become ever more intense. A large fraction of China’s emerging AI industry is focused on developing facial recognition and other surveillance technologies, and these companies are finding eager customers, not just in China but also in countries throughout the world. And, as we’ll see, AI-based surveillance technologies are by no means limited to authoritarian regimes. Facial recognition, in particular, is being widely deployed in the United States and other democratic countries and has already led to intense debate and accusations of bias and misuse. These issues will become only more fraught as the technology continues to become more powerful and—unless it is strictly regulated—ubiquitous.

中国跃居人工智能研发前沿

CHINA’S LEAP TO THE FOREFRONT OF ARTIFICIAL INTELLIGENCE RESEARCH AND DEVELOPMENT

2018 年 6 月，一场关于计算机视觉的大型会议在犹他州盐湖城举行。自 2012 年著名的 ImageNet 竞赛以来的六年里，机器视觉领域取得了长足的进步，研究人员现在专注于解决更为困难的问题。会议的亮点之一是“稳健视觉挑战赛”。这项由苹果和谷歌等大公司赞助的比赛让来自世界各地的大学和研究实验室的团队在一系列挑战中相互竞争，这些挑战旨在在不同情况下（例如室内或室外照明或不同的天气条件）可靠地识别图像。3^{这种能力对于自动驾驶汽车或在不同环境中运行的机器人等应用至关重要。比赛最重要的部分} ^之一集中在立体机器视觉上：使用两个摄像头，就像我们使用眼睛一样。通过从略有不同的有利位置解释视觉信息，我们的大脑能够生成场景的三维表示。两个正确放置的摄像头允许算法执行类似操作。4

In June 2018, a major conference on computer vision was held in Salt Lake City, Utah. In the six years since the famous 2012 ImageNet competition, the field of machine vision had advanced dramatically, and researchers were now focused on solving far more difficult problems. One of the highlights of the conference was the Robust Vision Challenge. This competition, sponsored by major companies including Apple and Google, pitted teams from universities and research labs across the world against each other in a series of challenges geared toward reliably identifying images in varying situations, such as indoor or outdoor lighting or differing weather conditions.³ This capability is crucial to applications like self-driving cars or robots that operate in varied environments. One of the most important segments of the contest focused on stereo machine vision: using two cameras in much the way that we employ our eyes. By interpreting visual information from slightly different vantage points, our brains are able to generate a three-dimensional representation of a scene. Two properly positioned cameras allow an algorithm to do something similar.⁴

获胜团队让很多人大吃一惊：他们是来自中国国防科技大学的研究人员。这所大学成立于 1953 年，前身是中国人民解放军军事工程学院，曾多次获得国家级研究和创新奖项，尤其是在计算机科学领域。据该大学网站介绍，该大学“以党的创新理论为基础开展教育工作，培养忠诚合格的接班人”。5^这似乎很好地表明，中国的学术或商业人工智能研究与该国的政治、军事和安全机构之间存在着一条非常模糊的分界线。

The winning team took many people by surprise: a group of researchers from China’s National Defense Technology University. The university was founded in 1953 as the People’s Liberation Army (PLA) military academy of engineering and has received numerous national awards for research and innovation, especially in computer science. According to its website, the university “bases its educational efforts on the Party’s innovation theory to cultivate loyal and qualified successors.”⁵ That seems like a pretty good indication that there is, at best, a very porous dividing line between academic or commercial AI research in China and the country’s political, military and security apparatus.

当然，中国政府干预和控制国内经济和社会的几乎每个方面都是常态。然而，中国最近在人工智能领域的快速发展，却受到中央政府明确产业政策的极大推动和推动。

It is, of course, routine for the Chinese government to intervene in and exert some degree of control over nearly every aspect of the country’s economy and society. However, China’s recent rapid progress in artificial intelligence has been significantly accelerated and orchestrated by an explicit industrial policy articulated by the central government.

许多观察家认为，中国共产党对人工智能兴趣突然高涨的催化剂是 2016 年 3 月 DeepMind 的 AlphaGo 系统与围棋冠军李世石之间的比赛，这场比赛备受瞩目。围棋起源于至少 2500 年前的中国，在中国公众中广受欢迎和推崇。AlphaGo 以 4-1 的比分在韩国首尔举行了为期七天的比赛，中国有超过 2.8 亿人观看了现场直播，几乎是观看几个小时观看普通超级碗比赛的观众的三倍。计算机在如此根植于中国历史和文化的智力竞赛中击败顶级人类选手的景象给公众以及中国学者、技术专家和政府官员留下了不可磨灭的印象。北京风险投资家兼作家李开复将 AlphaGo 与李世石的比赛称为“中国的斯普特尼克时刻”。斯普特尼克是苏联在 20 世纪 50 年代发射的一颗卫星，激发了公众对美国太空计划的支持^。6

Many observers believe that the catalyst for the sudden surge of interest in AI on the part of the Chinese Communist Party was the highly touted contest between DeepMind’s AlphaGo system and Go champion Lee Sedol that took place in March 2016. The game of Go originated in China at least 2,500 years ago and is wildly popular and revered among the Chinese public. AlphaGo’s 4–1 triumph, which took place over seven days in Seoul, South Korea, was viewed live by more than 280 million people in China—nearly three times the audience that tunes in for a few hours to watch a typical Super Bowl. The specter of a computer defeating a top human player at an intellectual pursuit so deeply rooted in Chinese history and culture made an indelible impression on the public as well as on Chinese academics, technologists and government bureaucrats. Kai-Fu Lee, a Beijing-based venture capitalist and author, calls the AlphaGo–Lee Sedol match “China’s Sputnik moment,” in reference to the Soviet satellite that galvanized public support for the U.S. space program in the 1950s.⁶

仅仅一年多之后，第二届比赛在中国乌镇举行。在三局比赛中，AlphaGo 以三连胜击败了当时世界排名第一的中国选手柯洁，获胜奖金为 150 万美元。不过，这一次没有现场观众。中国政府可能已经预料到了结果，发布了审查令，禁止对比赛进行任何现场直播，甚至禁止现场文字评论^。7

Just over a year later, a second contest was held in Wuzhen, China. In a three-game match carrying a $1.5 million prize for the winner, AlphaGo defeated the Chinese player Ke Jie, who was then ranked number one in the world, by prevailing in three straight games. This time around, however, there was no live audience. The Chinese government, perhaps anticipating the outcome, had issued a censorship order forbidding any live broadcast, or even live text-based commentary, of the match.⁷

2017 年 7 月，在柯洁输给 AlphaGo 两个月后，中国政府发布了一项明确的计划，将人工智能列为国家战略重点。这份名为《新一代人工智能发展规划》的文件宣称，人工智能将“深刻改变人类社会生活、改变世界”，并制定了到 2030 年实现人工智能技术主导地位的宏伟目标。该计划的作者写道，到 2020 年，中国“人工智能整体技术和应用与世界先进水平同步”，“人工智能产业成为新的重要的经济增长点”。接下来，“到 2025 年，中国将在人工智能基础理论方面取得重大突破，一些技术和应用达到世界领先水平，人工智能成为中国产业升级和经济转型的主要驱动力。”最后，“到2030年，我国人工智能理论、技术和应用达到世界领先水平，成为世界主要的人工智能创新中心，智能经济、智能社会应用取得明显成效，为进入创新型国家和经济强国行列奠定重要基础。” ⁱ , ⁸

Two months after Ke Jie’s loss to AlphaGo, in July 2017, the Chinese central government released an explicit plan designating artificial intelligence as a strategic national priority. Entitled “New Generation Artificial Intelligence Development Plan,” the document declared that AI was poised to “profoundly change human society and life and change the world” and then went on to lay out an extraordinarily ambitious step-by-step course toward domination of the technology by the year 2030. By 2020, the plan’s authors wrote, China’s “overall technology and application of AI will be in step with globally advanced levels” and “the AI industry will have become a new important economic growth point.” Next, “by 2025, China will achieve major breakthroughs in basic theories for AI, such that some technologies and applications achieve a world-leading level and AI becomes the main driving force for China’s industrial upgrading and economic transformation.” And finally, “by 2030, China’s AI theories, technologies, and applications should achieve world-leading levels, making China the world’s primary AI innovation center, achieving visible results in intelligent economy and intelligent society applications, and laying an important foundation for becoming a leading innovation-style nation and an economic power.”ⁱ,⁸

该文件的发布至关重要，不仅因为中国中央政府有能力直接微观管理全国人工智能能力的发展，还因为它确定了总体战略，或许更重要的是，为地区和地方政府提供了明确的激励机制。在中国的体制下，大量权力被下放给管理全国各个地区和城市的共产党官员。党内晋升主要靠任人唯贤，官员的职业发展轨迹在很大程度上取决于他或她在竞争激烈的生态圈中的表现，而这个生态圈重视以具体指标衡量的绩效。对于那些成功脱颖而出的人来说，可能几乎没有什么限制。习近平的大部分职业生涯都是在福建和浙江两省担任高级官员，后来又在上海市任职。

The publication of this document was crucial not because the Chinese central government has the ability to directly micromanage the development of artificial intelligence capability throughout the country, but because it defined an overall strategy and, perhaps more importantly, created clear incentives for regional and local governments. In China’s system, a great deal of power is delegated to the Communist Party officials who run the country’s various regions and cities. Advancement through the party’s ranks is largely meritocratic, and an official’s career trajectory is heavily dependent on how he or she performs in a competitive ecosphere that values performance as measured by specific metrics. For those who manage to stand out, there are potentially few limits. Xi Jinping spent much of his career as a top official in the Fujian and Zhejiang provinces and later in the city of Shanghai.

早在中央政府明确支持人工智能之前，特定地区就已开始投入大量资金，鼓励人工智能初创公司的发展。这些投资大多集中在高科技走廊，如南方城市深圳和北京西北部的中关村地区。中关村地区靠近中国两所最负盛名的大学——北大和清华，常被称为“中国硅谷”。然而，2017 年发布的战略文件实际上创造了一个明确的人工智能指标，地方官员知道他们很可能会根据这个指标接受评判。因此，全国各地的地区和城市迅速加入到竞争中，创建了经济特区和创业孵化器，并向人工智能初创公司提供直接的风险投资和租金补贴。单个城市的投资很容易达到数十亿美元。这种以创新为重点、协调松散的自上而下的指令在美国是难以想象的。美国版的地区间竞争通常是一种零和现象——德克萨斯州吸引加利福尼亚州的企业，或者城市为大公司提供大幅税收减免，以换取创造就业机会的设施。

Even before the central government explicitly embraced artificial intelligence, specific regions were already making substantial investments and encouraging AI startup companies. Much of this was focused in high-tech corridors such as the southern city Shenzhen and the Zhongguancun area in northwestern Beijing, which is close to the country’s two most prestigious universities, Peking and Tsinghua, and is often referred to as “China’s Silicon Valley.” However, publication of the strategy document in 2017 effectively created an explicit AI metric upon which regional officials knew they would likely be judged. As a result, regions and cities across the country quickly jumped into the fray, creating special economic zones and startup incubators and providing direct venture capital and rent subsidies to AI startups. The investments made by a single city can easily reach billions of dollars. This kind of loosely coordinated top-down directive with a focus on innovation would be hard to imagine in the United States. The American version of interregional competition is generally a far more zero-sum phenomenon—Texas luring businesses from California or cities offering large tax breaks to big companies in return for job-creating facilities.

中国在人工智能领域取得的进展具有许多关键优势。这些优势中的许多直接源于该国庞大的人口。截至 2020 年 3 月，中国拥有约 9 亿活跃互联网用户，超过美国和欧洲的总和，约占全球在线总人数的五分之一。9^然而 ^，互联网接入仅覆盖了约 65% 的人口，而美国则为 90%。10换句话说，中国的在线增长潜力要大得多。在中国 14 亿人口中，有大量聪明而雄心勃勃的高中生和大学生，他们渴望精通深度学习等技术，希望最终加入或创办中国数量激增的人工智能初创企业之一，其中许多企业的估值已超过 10 亿美元。这些年轻人是美国顶尖大学（如麻省理工学院和斯坦福大学）提供的在线课程最专注、最热情的参与者。他们还热切地阅读北美和欧洲顶尖人工智能研究人员发表的技术论文。因此，中国正在迅速培养一大批才华横溢、极其勤奋的工程师，他们不断吸收西方产生的最先进知识，并将很快准备将人工智能应用于中国经济和社会的几乎每个领域。

China enjoys a number of critical advantages as it powers forward in artificial intelligence. Many of these advantages derive directly from the country’s massive population. As of March 2020, China had about 900 million active internet users, more than the United States and Europe combined and roughly one fifth of the total number of people online globally.⁹ Internet access, however, had been extended to only about sixty-five percent of its population, versus ninety percent in the United States.¹⁰ In other words, China has far more online growth potential. Among China’s 1.4 billion people, there are an enormous number of smart and ambitious high school and university students who are eager to become proficient in technologies like deep learning in the hope of eventually joining—or launching—one of the exploding number of Chinese AI startups, many of which have achieved valuations over a billion dollars. These young people are among the most dedicated and enthusiastic participants in online courses offered by top U.S. universities like MIT and Stanford. They also eagerly comb through technical papers published by top AI researchers in North America and Europe. As a result, China is quickly developing a large pool of talented and extremely hard-working engineers who are continuously hoovering up state-of-the-art knowledge generated in the West and will soon be poised to leverage AI across virtually every dimension of China’s economy and society.

然而，最重要的优势在于中国经济活动产生的数据量和类型。作为一个发展中国家，中国在传统基础设施方面的投资少得多，因此，中国直接跃居移动技术的前沿。中国民众使用智能手机的活动范围比西方人普遍广泛得多。这一切尤其得益于腾讯微信应用的流行。微信于 2011 年推出，在中国以及其他国家华人中都广受欢迎。

The most important advantage, however, lies in the sheer volume, as well as the type, of data that Chinese economic activity generates. As a developing country, China has far less invested in legacy infrastructure, and as a result, the country has leapt directly to the very frontier of mobile technology. The Chinese public uses smartphones in a vastly wider array of activities than is typical in the West. All this is driven especially by the popularity of Tencent’s WeChat app. Introduced in 2011, WeChat has gained overwhelming popularity in China and also among the Chinese diaspora in other countries.

微信本质上是一款消息应用程序，与 Facebook 的 WhatsApp 大致相当。然而，腾讯很早就决定大幅扩展微信的功能，允许第三方使用所谓的“官方账户”添加自己的功能。这些账户本质上相当于小程序，在各种类型的企业中都非常受欢迎，尤其是与微信的数字支付功能相结合时。在美国和其他西方国家，每家企业都有自己的移动应用程序是常态。在中国，微信已经发展成为一种“主应用程序平台”，数百万企业和组织使用它与公众互动。中国人不仅用微信交流，还用它来支付餐厅账单、预约医生、在线约会、支付水电费、叫出租车，以及做几乎所有其他事情。通过微信提供的服务数量不断增加。与 Apple Pay 等需要商家投资昂贵的销售点设备的系统不同，微信移动支付只需显示条形码供客户扫描即可实现。因此，即使是街头食品摊贩等最小的企业也可以轻松接受数字支付。在中国，微信支付比信用卡更受欢迎，甚至在许多场所取代了现金。

At its core, WeChat is a messaging app roughly comparable to Facebook’s WhatsApp. However, Tencent made a decision early on to dramatically expand the capability of WeChat by allowing third parties to add their own functionality using what are called “official accounts.” These essentially amount to mini apps and are extraordinary popular with businesses of all types, especially when combined with WeChat’s ability to make digital payments. In the United States and in other western countries, the norm is for every business to have its own mobile app. In China, WeChat has evolved into a kind of “master app as platform,” and millions of businesses and organizations use it to interface with the public. The Chinese use WeChat not just to communicate, but also to pay bills at restaurants, to book doctor appointments, for online dating, to pay their utility bills, to hail taxis and, in essence, to do just about everything else. And the number of services available through WeChat expands continuously. Unlike systems such as Apple Pay, which require merchants to invest in expensive point-of-sale equipment, mobile payment via WeChat can be implemented simply by displaying a barcode for customers to scan. As a result, even the smallest businesses, such as street food vendors, can easily accept digital payments. Throughout China, WeChat payment is far more popular than credit cards and is even displacing cash in many venues.

结果是，中国的数字活动比美国多得多，而且这种活动对整体经济的影响也比美国深得多——捕捉到了大量在美国或欧洲可能离线进行的交易。每笔付款、每笔预订、每趟出租车和任何类型的互动都会产生数据，而这些数据非常适合深度学习算法处理。

The upshot is that there is vastly more digital activity in China, and it extends to a far greater depth in the overall economy—capturing a torrent of transactions that, in the United States or Europe, would likely be offline. And every payment, every booking, every taxi ride and every interaction of any type generates data that is ideally suited to be gobbled up by deep learning algorithms.

除了数据更加丰富之外，中国 AI 企业家通常也更容易获得数据。虽然确实存在数据隐私法规，但它们远不如美国或欧洲那样严格。公众往往也不会特别关注这些问题中的大多数。对个人隐私或算法中可能存在的种族偏见的担忧——这些问题可能会迅速在民主社会引起强烈愤慨——在中国不存在或几乎不会引起波澜。虽然谷歌最初与 DeepMind 签订合同获取 NHS 数据的行为立即在英国引发了强烈抗议，但在利用医疗和教育等领域的人工智能时，中国科技公司通常受益于更顺畅的实施和盈利途径。如果说数据是新的石油，那么中国的 AI 企业家就是新时代的野猫钻探者——在相对不受监管的数字领域中，在每个有希望的地方钻井和安装泵以提取价值。

In addition to being more abundant, data is also generally far more accessible to AI entrepreneurs in China. Though data privacy regulations do exist, they are nowhere near as stringent as in the United States or especially in Europe. Nor does the public tend to be particularly focused on most of these issues. Concerns about personal privacy, or perhaps possible racial bias in algorithms—issues that can quickly generate incandescent outrage in democratic societies—are nonexistent or barely cause a ripple in China. While Google’s access to NHS data that was originally contracted to DeepMind immediately led to an outcry in the United Kingdom, Chinese tech companies generally benefit from a smoother path to implementation and profitability when it comes to leveraging artificial intelligence in areas like healthcare and education. If data is the new oil, then China’s AI entrepreneurs are new age wildcatters—drilling and erecting pumps to extract value at every promising location across a relatively unpoliced digital terrain.

早在风险投资支持的人工智能初创企业蓬勃发展之前，中国的主要科技公司，尤其是腾讯、阿里巴巴和百度，就已经在人工智能研发方面投入了巨额资金。百度通常被称为“中国的谷歌”，是中国领先的互联网搜索引擎，在语音识别和语言翻译等领域拥有深厚的专业知识，但也在积极进军其他领域。例如，2017 年，百度推出了开源自动驾驶汽车平台 Apollo，本质上是一种“自动驾驶汽车的安卓”，该公司免费向中国高度分散的汽车制造业的公司提供该平台。包括宝马、福特和大众在内的^{11 家} 全球汽车公司以及 NVIDIA 等技术提供商也已签约成为合作伙伴。作为回报，百度可以访问汽车生成的数据，然后可以使用这些数据来训练其算法。换句话说，百度正在采取一种独特的战略，最终可能会提供与特斯拉数十万辆配备摄像头的汽车类似的优势。

Even before the explosion of venture-backed AI startups, China’s major technology companies, especially Tencent, Alibaba and Baidu, were making massive investments in artificial intelligence research and development. Baidu, which is often called “the Google of China” and is the country’s leading internet search engine, has developed deep expertise in areas like speech recognition and language translation, but is also pushing aggressively into other areas. In 2017, for example, Baidu introduced Apollo, an open-source autonomous vehicle platform—essentially a kind of “Android for self-driving cars”—which the company gives away freely to companies within China’s highly fragmented automotive manufacturing industry.¹¹ Global car companies, including BMW, Ford and Volkswagen, as well as technology providers like NVIDIA, have also signed on as partners. In return, Baidu gets access to the data that is generated by the vehicles, which it can then use to train its algorithms. In other words, Baidu is following a unique strategy that may eventually offer advantages similar to what Tesla enjoys with its hundreds of thousands of camera-equipped cars.

中国早期人工智能的发展很大程度上得益于美国和其他西方国家的知识和人才转移。精通中文的美国研究人员尤其成为百度的招聘目标。例如，2014 年，百度聘请了美国最知名的深度学习专家之一吴恩达 (Andrew Ng)，当时吴恩达正在领导谷歌大脑项目，这是谷歌首次利用大规模深度神经网络。吴恩达在百度工作了三年，之后返回硅谷，在北京建立了百度的主要人工智能研究实验室。2017 年，百度聘请了微软顶级人工智能高管陆奇担任公司首席运营官。^陆奇拥有卡内基梅隆大学的博士学位，他是越来越多在美国顶尖研究生课程接受教育的移民之一，他们选择返回中国，因为以人工智能为中心的商业机会往往被认为更具吸引力。事实上，大量的机会和快速变化的形势往往导致中国优秀人工智能专家的频繁离职。陆奇在百度只待了大约一年，现在在北京经营着一家初创企业孵化器。

Much of the early Chinese AI progress was driven significantly by transfers of knowledge and talent from the United States and other Western countries. American researchers with proficiency in the Chinese language, in particular, have been targeted for recruitment. In 2014, for example, Baidu hired one of the U.S.’s highest-profile deep learning experts, Andrew Ng, who was then leading the Google Brain project, the first initiative to leverage large-scale deep neural networks at Google. Ng, who stayed with the company for three years before returning to Silicon Valley, set up Baidu’s primary artificial intelligence research lab in Beijing. Then in 2017, Baidu hired Qi Lu, a top AI executive at Microsoft, to be the company’s chief operating officer.¹² Lu, who holds a PhD from Carnegie Mellon University, is one of a growing number of immigrants, educated at top American graduate programs, who are choosing to return to China because the business opportunities centered on AI are often perceived as more attractive. Indeed, the abundant opportunities and rapidly shifting terrain often lead to high turnover among talented Chinese AI experts. Lu remained with Baidu for only about a year and now runs a startup incubator in Beijing.

获得西方的研究成果和算法也发挥了关键作用。AlphaGo 击败柯洁大约一年后，腾讯宣布其自己的围棋软件 Fine Art 也成功击败了柯洁。然而，腾讯的系统很可能受到 DeepMind 已发表作品的很大启发，甚至可能直接抄袭自 DeepMind 已发表的作品。与我交谈过的大多数西方人工智能研究人员似乎并不特别关心这种知识转移，也不认为进步是国家竞争的结果；他们坚信一个强调公开发表研究成果和自由交流思想的全球体系。当我向 DeepMind 首席执行官 Demis Hassabis 询问所谓的“与中国的人工智能竞赛”时，他告诉我 DeepMind 公开发表论文，他知道“腾讯创造了一个 AlphaGo 克隆”，但他并不认为这是“这种意义上的竞赛，因为我们认识所有的研究人员，而且有很多合作。” ¹³

Access to research and algorithms developed in the West has also played a key role. About a year after AlphaGo defeated Ke Jie, Tencent announced that its own Go-playing software, Fine Art, had also succeeded in defeating the Go master. Tencent’s system, however, was likely heavily inspired by, or perhaps even directly copied from, DeepMind’s published work. Most of the Western AI researchers with whom I spoke don’t seem particularly concerned about this kind of knowledge transfer or view progress in terms of national competition; they believe strongly in a global system that emphasizes open publication of research and a free exchange of ideas. When I asked DeepMind CEO Demis Hassabis about a perceived “AI race with China,” he told me that DeepMind publishes openly and that he knows “Tencent has created an AlphaGo clone,” but that he doesn’t view it as “a race in that sense because we know all the researchers and there’s a lot of collaboration.”¹³

此外，据悉，中国研究人员为人工智能研究成果的发表做出了巨大贡献。根据艾伦人工智能研究所 2019 年初的一项分析，早在 2006 年，中国在人工智能研究论文发表总量方面就已超过美国。14^由于普遍认为这些论文中的许多质量相对较低或进展缓慢，艾伦研究所进一步分析了数量较少但被其他研究人员大量引用的论文。分析发现，假设目前的趋势继续下去，到 2019 年底，中国在论文发表量排名前 50%（按引用量衡量）方面将超过美国，到 2020 年，中国将在论文发表量排名前 10%（按引用量衡量）方面超过美国。到 2025 年，中国研究人员有望发表比美国更多真正优秀的论文（按引用量排名前 1%）。从另一个指标来看，中国在人工智能专利申请总数方面也已经领先于美国。

Also, by all accounts, Chinese researchers have been making substantial contributions to the published body of AI research. According to an early 2019 analysis by the Allen Institute for AI, China had already surpassed the United States in terms of total research papers on artificial intelligence published as far back as 2006.¹⁴ Because there is a general consensus that many of these papers are of relatively low quality or report very incremental progress, the Allen Institute did a further analysis focused on the smaller number of published papers that were heavily cited by other researchers. The analysis found that, assuming the continuation of current trends, China would have passed the United States in terms of publication of papers in the top fifty percent as measured by citations by the end of 2019 and the top ten percent most cited papers by 2020. Chinese researchers were on track to publish more truly elite papers—ranking in the top one percent in terms of citations—than the U.S. by 2025. By still another metric, China is also already ahead of the United States in terms of the total number of artificial intelligence patents filed.

但并非所有人都相信中国在人工智能研发方面即将超越美国。牛津大学人类未来研究所人工智能治理中心研究员 Jeffrey Ding 于 2018 年进行了一项分析，根据四个指标对美国和中国的人工智能能力进行了评级：人工智能计算硬件的安装基础、适用于机器学习的数据的可用性、研究和高级算法开发的能力以及商业人工智能生态圈的实力。基于这些因素，丁教授得出了他所谓的“人工智能潜力指数”，发现中国的评分仅为 17，而美国为 33。15^丁教授指出，例如，在中国发起的人工智能专利中，只有约 4% 后来也在其他司法管辖区提交了申请，这可能表明中国人工智能专利质量较低。 2019 年 6 月，他在向美国国会委员会作证时指出，中国在人工智能领域占据主导地位的说法被夸大了，美国仍然具有显著的结构性优势，美国的政策应该以维持现状为重点^。16

Not everyone buys into the idea that China is on the verge of surpassing the United States in artificial intelligence research and development. Jeffrey Ding, a researcher at the Centre for the Governance of AI at Oxford University’s Future of Humanity Institute, conducted an analysis in 2018 that rated AI capability in both the U.S. and China according to four metrics: the installed base of AI computing hardware, availability of data suitable for machine learning, proficiency in research and advanced algorithm development and the strength of the commercial AI ecosphere. Based on these factors, Ding derived what he calls an “AI Potential Index” and found that China rated only 17 versus 33 for the United States.¹⁵ Ding points out, for example, that only about four percent of AI patents initiated in China are later also filed in other jurisdictions, a likely indicator of low quality. In testimony before a U.S. congressional committee in June 2019, he argued that China’s purported rise to AI dominance has been overhyped, that the U.S. continues to have significant structural advantages and that American policy should be focused on maintaining the status quo.¹⁶

相比之下，李开复认为，美国可能仍将在人工智能前沿研究方面保持优势，但这一优势很快就会被中国在实际应用中将这项技术应用于整个经济领域的熟练程度所压倒。李开复认为，将人工智能应用于商业领域并不需要顶尖的、有远见的研究人员，而只需要大量能干、勤奋的工程师，以及能够轻松获取大量可用于训练机器学习算法的数据^。17

Kai-Fu Lee, in contrast, believes that the United States will likely continue to have an edge in research at the very frontier of artificial intelligence, but that this advantage will soon be overwhelmed by Chinese proficiency in doing the practical, nuts-and-bolts work of actually implementing the technology in applications across the economy. Lee argues that putting AI to work in the commercial sphere doesn’t require top-flight visionary researchers but simply large numbers of competent and diligent engineers with easy access to a deluge of data that can be used to train machine learning algorithms.¹⁷

人工智能的影响绝不会仅限于商业领域，这一显而易见的现实大大增加了中美之间任何人工智能竞赛的风险。人工智能将带来巨大的优势，可广泛应用于军事和国家安全领域。中国政府敏锐地意识到了这一点，并积极采取行动消除这两个领域之间的任何界限。2017 年，为响应习近平的直接倡议，中国宪法进行了修改，明确要求商业领域产生的任何技术进步必须与人民解放军共享。这就是所谓的“军民融合”原则。2018 年，百度与一家专注于电子战技术的中国军事研究所合作，为军方开发智能指挥和控制技术。负责该项目的百度高管尹世明是另一个在 SAP 和 Apple 等西方公司工作积累了丰富经验的工程师。在宣布合作的活动中，尹同表示，百度和军事学院将“携手合作，打通计算、数据和逻辑资源，进一步推进新一代人工智能技术在国防领域的应用。” ¹⁸

The stakes of any perceived AI race between the United States and China are raised greatly by the obvious reality that the impact of artificial intelligence will by no means be limited to the commercial sector. AI will deliver massive advantages that can be widely leveraged in military and national security applications. The Chinese government is keenly aware of this and has moved aggressively to erase any line between these two spheres. In 2017, in response to a direct initiative by Xi Jinping, the Chinese constitution was modified to explicitly require that any technological advances generated in the commercial sector must be shared with the People’s Liberation Army. This is known as the principle of “military-civil fusion.” In 2018, Baidu partnered with a Chinese military institute focused on electronic warfare technology on a project to develop intelligent command and control technology for the military. The Baidu executive in charge, Yin Shiming, was yet another example of an engineer who had developed deep experience working at Western companies, including SAP and Apple. At an event announcing the partnership, Yin declared that Baidu and the military institute would “work hand in hand to link up computing, data and logic resources to further advance the application of new generation AI technologies in the area of defense.”¹⁸

相比之下，不满的员工向谷歌施加压力，要求其停止竞标五角大楼的 JEDI 云计算合同。另一项国防计划 Maven 项目涉及开发可用于分析从美国军用无人机收集的图像的计算机视觉算法，这在谷歌员工中引起了更大的愤怒。2018 年，超过 3,000 名员工签署了一份反对该项目的请愿书，一些技术专家离开了公司。19与^JEDI一样，谷歌最终放弃了该项目。虽然谷歌员工当然有权表达自己的观点，但我认为这里的不对称既明显又令人不安。百度或腾讯的员工可以（或会）提出类似的抗议，坦率地说，这种想法是荒谬的。不可回避的事实是，民主国家公民享有的自由不是仅仅为了行使而存在的固有人权——相反，它们是面对威权主义必须捍卫的政治权利。在中美两国人工智能技术整体水平趋于平衡的情况下，如果谷歌等公司不愿与美国军方和安全机构合作，而中国公司却有义务协助中国的威权政权，而这一义务已经明确写入中国宪法，那么美国如何在国家安全基础上进行竞争？

Contrast this with the pressure that unhappy employees put on Google to end its bid to compete for the Pentagon’s JEDI cloud computing contract. Another defense initiative, Project Maven, which involved the development of computer vision algorithms that could be used to analyze images collected from U.S. military drones, generated even more outrage among Google workers. In 2018, more than 3,000 employees signed a petition objecting to the project, and a number of technical experts left the company.¹⁹ As with JEDI, Google ultimately abandoned the project. Though Google employees certainly have every right to express their views, the asymmetry here is, I think, both obvious and disturbing. The idea that workers at Baidu or Tencent could (or would) lodge a comparable protest is, frankly, absurd. There’s no getting around the fact that the freedoms enjoyed by the citizens of democratic countries are not intrinsic human rights that exist simply to be exercised—rather, they are political rights that have to be defended in the face of authoritarianism. As overall proficiency in artificial intelligence technologies approaches parity between the two countries, how can the United States compete on a national security footing if companies like Google are reluctant to cooperate with American military and security agencies, while their Chinese counterparts face an obligation to assist China’s authoritarian regime so explicit that it is written into the country’s constitution?

在我看来，美国和其他西方国家显然需要认真对待中国在人工智能领域的快速崛起。这可能需要政府加大对大学基础研究的支持。同样重要的是，美国必须继续利用其最重要的优势之一：其大学和科技公司一直吸引着来自世界各地的人才。我在 2018 年出版的《智能建筑师》一书中采访了 23 位顶尖人工智能研究人员，他们的背景清楚地表明了美国需要对高技能移民持开放态度。在我采访的 23 人中，目前有 19 人在美国工作。然而，在这 19 人中，超过一半出生在美国以外。原籍国包括澳大利亚、中国、埃及、法国、以色列、罗得西亚（现津巴布韦）、罗马尼亚和英国。如果美国不能继续吸引来自世界各地的最聪明的计算机科学家，中国将不可避免地获得优势，因为它将继续加大对人口教育的投资，而人口数量大约是美国的四倍。

To me, it seems clear that the United States and other Western countries need to take China’s rapid rise in artificial intelligence very seriously. This will likely call for increased government support for basic research at universities. It’s also critical that the U.S., in particular, continues to leverage one of its most important advantages: the fact that its universities and technology companies have been a magnet for talent from around the globe. The need for U.S. openness to high-skill immigration is clearly demonstrated by the backgrounds of the twenty-three top AI researchers I interviewed for my 2018 book Architects of Intelligence. Nineteen of the twenty-three people I spoke with currently work in the U.S. Of those nineteen, however, more than half were born outside the United States. Countries of origin include Australia, China, Egypt, France, Israel, Rhodesia (now Zimbabwe), Romania, and the United Kingdom. If the United States fails to continue attracting the brightest computer scientists from around the world, China will inevitably gain an advantage as it continues to invest more in educating a population that is roughly four times that of the United States.

中国监控国家的崛起

THE RISE OF CHINA’S SURVEILLANCE STATE

中国的威权政府体制与其创业型人工智能生态圈之间的强大协同作用在人脸识别技术初创企业群体的爆炸式增长中表现得最为明显。截至 2020 年初，该群体中的四家公司——商汤科技、云从科技、旷视科技和依图科技——已获得“独角兽”地位，即市值超过 10 亿美元。20虽然^分析人士可能会争论中国在整体人工智能技术方面是否已接近美国，但毫无疑问，在用于分析和识别人脸及其他属性的深度学习算法方面，中国公司处于绝对领先地位。与中国正在部署人工智能的其他领域一样，所有这些进步的关键驱动力是获取可用于训练机器学习算法的海量数据。截至 2020 年，中国全国共安装了约 3 亿个监控摄像头，在从各种可能的情况和角度获取人脸数字照片方面，中国在全球遥遥领先。

NOWHERE IS THE powerful synergy between China’s authoritarian system of government and its entrepreneurial artificial intelligence ecosphere more evident than in the exploding cluster of startup companies focused on facial recognition technology. As of early 2020, four companies in this group—SenseTime, CloudWalk, Megvii and Yitu—had achieved “unicorn” status, or a market valuation of more than a billion dollars.²⁰ While analysts may debate whether China has reached something close to parity with the United States in the overall technology of artificial intelligence, there is little doubt that when it comes to deep learning algorithms deployed to analyze and recognize human faces and other attributes, Chinese companies are at the absolute forefront of the field. As with other areas where artificial intelligence is being deployed in China, a critical driver of all this progress is access to a massive deluge of data that can be used to train machine learning algorithms. With an estimated 300 million surveillance cameras installed throughout the country as of 2020, China, when it comes to the availability of digital photographs of human faces in every conceivable situation and from every possible angle, is far and away the global leader.

面部识别初创公司受到中国威权国家各个层面对监控技术似乎无限需求的推动。最热切的购买者之一是地方公安部门，他们越来越多地在自己的地区建立压迫性监控网络。虽然新疆仍然是中国监控国家的重灾区，但在那里测试和完善的技术正在迅速传播到全国各地。公安部门经常将面部识别系统与其他技术相结合，例如手机扫描仪（可捕获经过附近的每部手机的唯一识别码）、汽车牌照阅读器和指纹识别技术，以编织一幅奥威尔式的挂毯，随着时间的推移，这种挂毯变得越来越一体化。例如，算法通常可以将手机识别码与人脸匹配，从而为个人创建一个全面的跟踪和识别系统。此类系统安装在已知与较高犯罪率相关的社区或特定建筑物的入口处。进入住宅区也通常通过面部识别系统来实现，而不是通过钥匙卡或其他侵入性较小的方法。这使建筑管理和当地警察部门能够追踪居民和客人，并防止公寓的非法转租^。21

Facial recognition startups are buoyed by seemingly limitless demand for surveillance technology from every level of China’s authoritarian state. Some of the most eager purchasers of the technology are local police departments, who are increasingly setting up oppressive surveillance networks specific to their regions. While Xinjiang remains ground zero for the Chinese surveillance state, the technologies tested and perfected there are rapidly spreading across the country. Police departments often combine facial recognition systems with other technologies such as mobile phone scanners, which capture a unique identification code for every phone that passes through the vicinity, car license plate readers and fingerprint recognition technology to weave an Orwellian tapestry that is becoming increasing integrated over time. Algorithms can often, for example, match phone identification codes with faces, creating a comprehensive tracking and identification system for individuals. Such systems are installed in neighborhoods or at the entrances of specific buildings known to be associated with higher crime levels. Entry to housing complexes is also often enabled by facial recognition systems rather than by key cards or other less intrusive methods. This enables building management and local police departments to track residents and guests and also prevent illegal subletting of apartments.²¹

监控摄像头也密集分布在旅客经常光顾或人群可能聚集的任何区域，如火车站、体育场、旅游景点和活动场所。在一些备受关注的案件中，警方在多达 60,000 人的音乐会或节日上逮捕了特定的个人，而这完全是基于向当局发出面部识别匹配警报的算法。22在一个看似直接来自反乌托邦科幻电影的场景中，警方可以通过佩戴实验^性的面部识别眼镜来逮捕嫌疑人，只要目标保持静止几秒钟并被纳入区域面部识别数据库，这种眼镜就可以生成身份证明。其他人工智能系统可以根据人们所穿的衣服，甚至通过分析他们步态的独特特征来追踪他们。

Surveillance cameras are also heavily clustered in any areas visited by travelers or where crowds are likely to congregate, such as train stations, stadiums, tourist attractions and event venues. In a number of highly publicized cases, police have arrested specific individuals at concerts or festivals attended by as many as 60,000 people, purely on the basis of algorithms alerting authorities to a facial recognition match.²² In a scene seemingly straight out of a dystopian science fiction movie, police can apprehend suspects by wearing experimental facial recognition glasses that can generate an identification as long as the target remains still for several seconds and is included in a regional facial recognition database. Other AI systems can track people based on the clothing they are wearing or even through analyzing the unique characteristics of their gait.

该技术最著名的用途之一是襄阳市在繁忙的十字路口设置了一套系统，用于抓捕并羞辱闯红灯的人。该系统会捕捉非法穿越马路的人的照片，然后将这些照片与他们的身份进行匹配，并显示在大屏幕上，试图让他们受到公众的羞辱和八卦。23^在包括上海在内的其他城市，也有类似的系统开罚单。可以肯定的是，在中国，面部识别技术并不是所有用途都专门用于监视。在零售店授权付款、购买火车票或登机方面，中国处于领先地位。然而，通过常规使用该技术产生的任何数据几乎肯定会被警察部门和安全机构获取。

In one of the most famous uses of this technology, the city of Xiangyang set up a system to catch and then embarrass jaywalkers at a busy intersection. The system captures photos of people crossing the street illegally that are then later matched with their identities and displayed on a large screen in an attempt to subject them to public shaming and gossip.²³ In other cities, including Shanghai, similar systems issue fines. To be sure, not all uses of facial recognition technology in China are specifically geared toward surveillance. The country is a leader in scanning faces to authorize payments at retail stores, to purchase train tickets or to board aircraft. However, any of the data generated through routine use of the technology is almost certainly available to police departments and security agencies.

虽然中国普遍存在的监控至少在一定程度上可以作为一种保护社会免受已知犯罪背景个人侵害的机制，但在其他情况下，它严重超越了道德界限，这在西方是不可想象的。例如，一些警察部门提出了具体要求，要求技术配置为不识别个人面孔，而是识别维吾尔族或其他“敏感人群”的种族特征。中国面部识别初创公司迅速采取行动，满足市场需求。2019 年 4 月，《纽约时报》的保罗·莫泽尔 (Paul Mozer) 发表了一篇文章，其中附有一张云从科技在线营销材料的截图，该材料承诺其技术的潜在买家，如果“社区中敏感人群的数量增加（例如，如果最初一个维吾尔族人住在一个社区，20 天内出现了六个维吾尔族人），它会立即发出警报，以便执法人员做出反应，询问这些人，处理情况，并制定应急计划。” ²⁴

While much of the pervasive surveillance in China can, at least to some extent, be defended as a mechanism to protect society from individuals with known criminal backgrounds, in other cases it violently transgresses ethical boundaries in ways that would be unthinkable in the West. Some police departments, for example, have put out specific requests for technology configured not to recognize individual faces, but rather the racial characteristics of Uyghurs or other “sensitive peoples.” Chinese facial recognition startups have moved rapidly to supply market demand. An April 2019 article by Paul Mozer of the New York Times included a screenshot of online marketing material from CloudWalk that promised potential buyers of its technology that if “the number of sensitive groups of people in the neighborhood increases (for example, if originally one Uyghur lives in a neighborhood, and within 20 days six Uyghur appear), it immediately sends alarms so that law enforcement personnel can respond, question the people and handle the situation, and develop a contingency plan.”²⁴

除了显示维吾尔族家庭和平站立、维吾尔族人排队从武警身边走过以及民众骚乱场景的图像外，云从科技的网站在“社区管控和敏感人群防范”标题下继续解释说，“在社区中，人脸识别系统收集这些人的身份和面部数据，同时火眼大数据平台收集敏感人群的身份、进出时间、人数等，并向警方发出警告，以便警方实现管控敏感人群的目标。” ²⁵

Next to images that appear to show a Uyghur family standing peacefully, a line of Uyghurs filing past militarized police and scenes of civil unrest, CloudWalk’s website goes on to explain under the heading “Neighborhood control and prevention of sensitive peoples” that “in the neighborhood, the facial recognition system collects these people’s identity and facial data, at the same time the Fire Eye big data platform collects sensitive groups’ identities, times of entry and exit, the number of individuals, etc., and issues warnings to police so they can carry out their goal of managing and controlling sensitive groups.”²⁵

甚至连 CloudWalk 产品的品牌“Fire Eye 大数据平台”都像是直接从一本特别可怕的科幻小说中剥离出来的。即使是在公开的公司网站上，在描述这项技术的意图时也完全没有采取任何诡计或微妙的手段，这相当戏剧性地表明了中国政府对维吾尔族人的压迫有多么明显，以及人工智能在坏人手中可以如何以真正反乌托邦的方式被利用。这里的危险绝不限于中国。几乎任何先进的面部识别技术都可能被用作对付特定群体的武器，只要配置系统来识别种族、性别、面部毛发或宗教服饰等属性。

Even the branding of CloudWalk’s product “Fire Eye big data platform” seems stripped directly from an especially frightening science fiction novel. The total lack of any attempt at subterfuge, or subtlety, in describing the technology’s intent, even on a publicly available corporate website, is a pretty dramatic indication of just how overtly oppressive the Chinese government’s campaign against the Uyghurs is, and how artificial intelligence, in the wrong set of hands, can be leveraged in genuinely dystopian ways. The danger here is by no means limited to China. Nearly any advanced facial recognition technology could conceivably be weaponized against specific groups by configuring the system to identify attributes such as race, gender, facial hair or religious attire.

^{中国对公民实施更全面监控的轨迹，可能随着该国计划中的社会信用体系的全面实施而达到顶峰。该}体系于 2014 年宣布实施，旨在奖励全民“守信”，宣称其意图是“让守信者天下无敌，让失信者寸步难行”。26社会信用体系始于西方商业信用或消费者评级体系的典型措施，例如根据个人还债历史进行评估，或类似 Uber 或 Airbnb 等服务使用的评级系统。但中国的体系走得更远，可能侵入日常生活的几乎每个方面，将违法行为以及任何被国家视为不良的行为都考虑在内。除了未能及时支付账单或罚款之外，这还可能包括玩太多电子游戏、在社交媒体上发布有争议的想法、与坏人交往、在公共交通工具上吃东西、乱扔垃圾或播放嘈杂的音乐、在禁止吸烟的地方吸烟，甚至未能正确分类垃圾。27社会信用计算还可以奖励积极行为，例如赢得公民或员工奖、向慈善机构捐款或为照顾家人或帮助邻居做出巨大努力。该系统甚至^可以深入到最私密的消费者决策，例如奖励被认为是积极的购买，如婴儿尿布，同时惩罚过量购买酒精。获得高分的人会获得额外奖励，如取暖费折扣、在医院或政府机构的等候时间更短，或优先获得最好的就业机会。另一方面，社会信用评分较低的人将面临惩罚，例如无法预订飞机和火车票、无法让孩子进入最好的学校或无法预订理想的酒店或度假村。一旦这样一个全面的系统全面投入运行，它将成为一种极其侵入性的控制机制，几乎持续施加于中国庞大人口中的每个成年人。人权观察组织恰当地称这个想法“令人不寒而栗” ^。28

China’s trajectory toward ever more comprehensive surveillance of its citizens may culminate with the full implementation of the country’s planned social credit system. Announced in 2014 as a way to reward “trustworthiness” throughout the population, the program’s declared intent is to “allow the trustworthy to roam everywhere under heaven while making it hard for the discredited to take a single step.”²⁶ The social credit system begins with measures that are typical of commercially administered credit or consumer rating systems in the West, such as evaluations based on a person’s history of paying debt obligations or the kind of rating systems used on services like Uber or Airbnb. But the Chinese system goes much further, potentially intruding into virtually every aspect of daily life by taking into account violations of the law, as well as any behaviors deemed undesirable by the state. In addition to failing to pay your bills or fines in a timely manner, this might include playing too many video games, posting controversial thoughts on social media, associating with the wrong people, eating, littering or playing loud music on public transit, smoking where it is prohibited or even failing to properly sort garbage.²⁷ The social credit calculation can also reward positive behaviors, such as winning a civic or employee award, giving money to charity or making an outsized effort to take care of family members or assist neighbors. The system can reach even into the most intimate consumer decisions, for example, rewarding purchases deemed positive, such as baby diapers, while penalizing excessive purchases of alcohol. Those who achieve superior scores are rewarded with perks such as discounts on heating bills, shorter wait times at hospitals or government agencies, or preferred access to the best employment opportunities. Those who, on the other hand, have low social credit scores face penalties such as an inability to book tickets on planes and trains, lack of access to the best schools for their children or blocks on attempts to make reservations at desirable hotels or resorts. Once such a comprehensive system is fully operational, it will be an extraordinary intrusive mechanism of control, exerted continuously on virtually every adult in China’s massive population. It’s an idea that Human Rights Watch has appropriately called “chilling.”²⁸

虽然所有这些都代表了最终的愿景，但当前的现实却远没有那么紧密。实际上，社会信用体系被分割成由各个城市和地方政府运行的实验项目，以及由阿里巴巴或腾讯等维护移动支付系统的公司管理的一系列商业评级系统。29^其中一些项目，例如荣成市的项目，得到了公众的广泛认可，因为它们相对透明，只惩罚明显违法的行为，并且产生了不可否认的积极效果。例如，荣成的司机在明确违规行为会对社会信用评级产生负面影响后，开始在人行横道上停车以尊重行人。数百万张机票或高铁票确实被拒绝发给中国公民，但这通常是因为他们的名字出现在长期使用的黑名单上，而不是算法生成的分数的结果。最重要的黑名单由最高人民法院保存，主要包括未偿还债务、法院判决或罚款的人——尽管与中国几乎所有政府职能一样，腐败和缺乏透明度是一个持续存在的问题。随着时间的推移，这些系统似乎不可避免地会变得更加一体化，面部识别和其他用于跟踪和监控公民的人工智能技术将放大它们的侵入性。最终，一个真正奥威尔式的全面、精心策划的社会控制系统很可能会出现。

While all that represents the ultimate vision, the current reality is far less cohesive. In practice, the social credit system is fragmented into experimental programs run by various city and local governments together with an array of commercial rating systems administered by corporations, like Alibaba or Tencent, that maintain mobile payment systems.²⁹ Some of the programs, such as one in the city of Rongcheng, have met with widespread approval from the public because they are relatively transparent, penalize only behaviors that are clearly illegal and have produced undeniably positive results. Drivers in Rongcheng, for example, began stopping at crosswalks in deference to pedestrians once it was made clear that violations would negatively impact their social credit ratings. Millions of airline or high-speed rail tickets have indeed been denied to Chinese citizens, but this is generally because their names have appeared on long-used blacklists rather than the result of an algorithmically generated score. The most important blacklist, maintained by the Supreme People’s Court, primarily includes people who have unpaid debts, court judgments or fines—although, as with nearly all government functions in China, corruption and a lack of transparency are a constant problem. Over time, it seems inevitable that these systems will become far more integrated, and their intrusiveness will be amplified by facial recognition and other AI technologies used to track and monitor citizens. Eventually, a genuinely Orwellian system of comprehensive and carefully orchestrated social control could well emerge.

这一切并不局限于中国。事实上，监控技术的出口在中国政府的总体战略中发挥着关键作用，该战略旨在将中国的生产从低利润商品转向高价值的技术产品。中国控制着全球近一半的人脸识别技术市场。其中大部分由一家中国公司主导——电信公司华为。根据卡内基国际和平基金会 2019 年 9 月的分析，华为已经向至少 50 个国家和 230 个城市出售了包括人脸识别在内的监控技术，远远超过任何其他单一公司。相比之下，最接近的美国竞争对手 IBM、Palantir 和思科都只在不到 12 个国家安装了系统。^{沙特阿拉伯} 和阿联酋等拥有威权政府的国家是中国技术尤其热切的客户，因为它们正在扩大自己的全国性监控系统。在这些国家，人脸识别往往已经成为日常生活的一部分。 2019 年初，我在阿布扎比旅行时了解到了这一点，当时我听到一个广为流传的故事，说一位富婆丢失了一枚昂贵的戒指。她向有关部门报告了这一不幸事件，有关部门立即将面部识别软件应用于相关区域的监控录像，并在事发后数小时内赶到了捡到戒指的人家门口。

None of this is limited to China itself. Indeed, the export of surveillance technology plays a key role in the Chinese government’s overall strategy to transform the country’s production away from low-margin commodities and toward higher-value technology products. China controls nearly half of the global market for facial recognition technology. Much of this is led by a single Chinese firm, the telecommunications company Huawei. According to a September 2019 analysis by the Carnegie Endowment for International Peace, Huawei has sold surveillance technology, including facial recognition, to at least fifty countries and 230 cities, far more than any other single company. By comparison, the closest U.S. competitors—IBM, Palantir and Cisco—have each installed systems in less than a dozen countries.³⁰ Nations with authoritarian governments, such as Saudi Arabia and the United Arab Emirates, are especially eager customers for Chinese technology as they expand their own countrywide surveillance systems. In these nations, facial recognition is often already a routine aspect of daily life. I learned this myself on a trip to Abu Dhabi in early 2019 where I heard a widely circulated story about a wealthy woman who lost an expensive ring. She reported the mishap to authorities, who immediately applied facial recognition software to surveillance footage of the relevant area—and arrived at the doorstep of the person who had picked up the ring within hours of the incident.

华为的监控设备销售通常由中国政府贷款资助。肯尼亚、老挝、蒙古、乌干达、乌兹别克斯坦和津巴布韦等国都参与其中，有些国家是北京“一带一路”倡议的一部分，该倡议为近 70 个国家的基础设施提供资金。非洲是一个越来越重要的焦点，据一些人称，中国的面部识别系统已经产生了重大影响。例如，华为声称，在肯尼亚首都内罗毕及其周边地区安装其技术使 2015 年犯罪率下降了 46% ^。31

Huawei’s sales of surveillance equipment are often financed by loans backed by the Chinese government. Countries including Kenya, Laos, Mongolia, Uganda, Uzbekistan and Zimbabwe have participated, in some cases as part of Beijing’s global Belt and Road Initiative, which is funding infrastructure in nearly seventy countries. Africa is an increasingly important focus, and by some accounts, Chinese facial recognition systems have already had a significant impact. Huawei claims, for example, that the installation of its technology in and around Kenya’s capital city of Nairobi led to a forty-six percent reduction in crime in 2015.³¹

中国公司正在开发的技术对安全和人权的影响已经导致与美国的重大摩擦。2019 年 5 月，华为受到贸易限制，导致美国禁止向该公司出售软件和计算机芯片等美国技术。这是由于美国在不断升级的全面贸易战中的某种姿态，以及长期以来人们的担忧，即该公司出售的 5G 手机基础设施技术如果在当地安装，可能会让中国政府获取美国通信。32 美国在向^盟国施加巨大压力以禁止使用华为设备方面也取得了好坏参半的成功。此外，华为被指控违反美国对伊朗的贸易禁运，并因接受中国政府的不当支持而受到指责。

The security and human rights implications of the technologies being developed by Chinese companies have already led to significant friction with the United States. In May 2019, Huawei was subjected to trade restrictions that resulted in a ban on the sale of U.S. technology such as software and computer chips to the company. This was driven by some combination of U.S. posturing amidst an escalating general trade war and long-held concerns that the 5G mobile phone infrastructure technology sold by the company could potentially allow the Chinese government to access communications in the United States if the equipment were installed locally.³² The U.S. also had mixed success in its effort to put significant pressure on allied countries to likewise prohibit use of Huawei’s equipment. In addition, Huawei was accused of violating the U.S. trade embargo against Iran and had been singled out for having received inappropriate support from the Chinese state.

五个月后，美国将贸易黑名单扩大到包括中国几家最重要的人工智能初创公司以及二十家中国警察部门或安全机构，表面上是因为这些公司将技术部署到维吾尔族和其他少数民族身上，导致侵犯人权。禁令包括中国四家面部识别独角兽中的三家，以及专门从事语音识别系统的科大讯飞，以及两家生产摄像头和其他监控硬件的公司^。33

Five months later, the United States extended the trade blacklist to include several of China’s most important artificial intelligence startups, as well as twenty Chinese police departments or security agencies, ostensibly because of human rights violations resulting from the deployment of their technology against the Uyghurs and other minority populations. Included in the ban were three of China’s four facial recognition unicorns, as well as iFlytek, which specializes in speech recognition systems, and two companies that manufacture cameras and other surveillance hardware.³³

在新冠疫情爆发后，中美之间的紧张关系显著升级，人们普遍认识到，过度依赖中国的生产可能会威胁到美国获取至关重要的战略材料以及医疗用品和药品的渠道。早在危机爆发之前，两国之间的经济协同和相互依存关系（历史学家尼尔·弗格森在 2006 年将这一现象称为“中美国”）显然正在逐渐减弱。如果紧张局势进一步加剧，两国继续脱钩，那么围绕人工智能开发和部署的冲突和竞争将不可避免地发挥核心作用，而且随着人工智能既是系统技术又是战略技术这一点越来越明显，两国之间全面爆发人工智能军备竞赛的可能性已成现实。

In the wake of the coronavirus pandemic, tensions between the United States and China have escalated significantly, and there is widespread recognition that overdependence on Chinese production potentially threatens U.S. access to critically important strategic materials, as well as healthcare supplies and medicines. Even before the crisis, it was clear that the economic synergy and interdependence between the two countries—a phenomenon that the historian Niall Ferguson dubbed “Chimerica” in 2006—was gradually unwinding. If tensions increase further and the two countries continue to decouple, it seems inevitable that conflict and competition centered on the development and deployment of artificial intelligence will play a central role, and as it becomes increasingly evident that AI is both a systemic and a strategic technology, the specter of a full-out AI arms race between the two countries looms as a genuine danger.

西方兴起的面部识别争论

AN EMERGING DEBATE OVER FACIAL RECOGNITION IN THE WEST

2019 年 2 月，印第安纳州警方正在调查一起发生在公园的犯罪事件。两名男子打架。其中一名男子掏出枪，向另一名男子的腹部开枪，然后逃离现场。一名旁观者用手机记录了这一事件，因此州警探决定尝试将袭击者的脸部图像上传到他们一直在试验的新型面部识别系统中。系统立即生成匹配结果；枪手出现在社交媒体上发布的视频中，并附有包括他的名字在内的描述。尽管嫌疑人之前从未被捕过，甚至没有驾照，但破案仅用了大约 20 分钟^。34

In February 2019, the Indiana State Police were investigating a crime that had occurred when two men got into a fight at a park. One of the men pulled out a gun, shot the other man in the abdomen and then fled the scene. A bystander had recorded the incident on a mobile phone, so the state police detectives decided to try uploading an image of the assailant’s face to a new facial recognition system with which they had been experimenting. The system generated a match immediately; the shooter had appeared in a video posted on social media along with a description that included his name. All told, it took about twenty minutes to solve the case, in spite of the fact that the suspect had not been arrested previously and did not even have a driver’s license.³⁴

侦探们通过一家名为 Clearview AI 的神秘公司提供的移动应用程序与面部识别系统进行交互。Clearview 应用程序可用的照片数据库非常庞大。该公司并不依赖与护照、驾照或面部照片相关的官方政府照片，而是简单地搜索互联网并从 Facebook、YouTube 和 Twitter 等各种来源抓取公开可用的图像。如果 Clearview 系统找到匹配项，该应用程序会显示匹配照片出现在网上的网页或社交媒体个人资料的链接——通常可以立即识别。Clearview 构建的数据集包括大约 30 亿张抓取的图像——是 FBI 维护的美国公民官方照片数据库的七倍多。这是一项了不起的成就，尤其是因为 Clearview AI 是一家小公司——比中国的面部识别独角兽小几个数量级——而且，至少在 2020 年 1 月之前，在执法圈外几乎完全不为人知^。35

The detectives interfaced with the facial recognition system through a mobile app provided by a mysterious company called Clearview AI. The database of photographs available to the Clearview app was truly massive. Rather than relying on official government photographs such as those associated with passports, driver’s licenses or mug shots, the company had simply scoured the internet and scraped publicly available images from a variety of sources including Facebook, YouTube and Twitter. If the Clearview system found a match, the app displayed links to the webpages or social media profiles where a matching photo had appeared online—often allowing for an immediate identification. The dataset constructed by Clearview included roughly three billion scraped images—more than seven times the size of the official photographic database of U.S. citizens maintained by the FBI. This was a remarkable feat, especially since Clearview AI was a tiny company—orders of magnitude smaller than the facial recognition unicorns in China—and, at least until January 2020, was almost entirely unknown outside of law enforcement circles.³⁵

当月，《纽约时报》刊登了科技记者克什米尔·希尔撰写的一篇重大调查报道，深入调查了该公司的背景，并首次曝光了其运营情况。事实证明，Clearview 在其 LinkedIn 页面上列出了一个不存在的纽约地址，它是由一位名叫 Hoan Ton-That 的澳大利亚连续创业者于 2016 年创立的。除了其他资金外，这家初创公司还从硅谷风险投资家彼得·泰尔那里获得了 20 万美元的种子资金，彼得·泰尔还共同创立了 Palantir，这是一家与安全机构和警察部门有密切联系的数据分析和监控公司。

That month the New York Times published a major investigative exposé by technology reporter Kashmir Hill that delved into the company’s background and, for the first time, shined a spotlight on its operations. It turned out that Clearview, which listed a nonexistent New York address on its LinkedIn page, had been founded in 2016 by an Australian serial entrepreneur named Hoan Ton-That. Among other funding, the startup had received $200,000 in seed money from Silicon Valley venture capitalist Peter Thiel, who had also co-founded Palantir, a data analytics and surveillance company with significant ties to security agencies and police departments.

Clearview 声称，它只向合法的执法机构或政府安全机构提供其技术。然而，理论上没有什么可以阻止该公司最终向公众开放其系统，这引发了几乎完全失去匿名性的担忧。一旦该技术得到广泛应用，几乎任何地方的任何人都可以被使用 Clearview 应用程序的随机陌生人立即识别。只要知道一个人的名字，就可以很容易地找到他的家庭住址、工作地点和各种其他敏感信息。不可避免的结果是，几乎任何不检点行为都会引发大量跟踪、勒索或公开羞辱，以及无数其他不当行为。换句话说，一个过度监控的反乌托邦——可能比中国正在考虑的任何事情都更具侵入性和可怕性——很可能在美国私营部门出现，而政府根本不参与或监督。Clearview AI 的一些支持者似乎并不特别担心这种可能性。 “我得出的结论是，由于信息不断增加，隐私将不复存在，”该公司的一位早期投资者告诉《纽约时报》。“法律必须确定什么是合法的，但你不能禁止技术。当然，这可能会导致反乌托邦的未来，但你不能禁止它。” ³⁶

Clearview claimed that it made its technology available only to legitimate law enforcement or government security agencies. However, there was nothing in theory to prevent the company from eventually making its system available to the public, raising the specter of a near complete loss of anonymity. Once the technology became widely available, virtually anyone, anywhere, could be instantly identified by a random stranger wielding Clearview’s app. Given a person’s name, it would be a simple matter to find a home address, place of employment and all kinds of other sensitive information. The inevitable outcome would be an explosion of stalking, blackmail or public shaming for virtually any indiscretion, and countless other misbehaviors. In other words, it seems quite possible that an overreaching surveillance dystopia—potentially more intrusive and frightening than anything being contemplated in China—could emerge from the private sector in the United States, without any government involvement or oversight at all. Some backers of Clearview AI didn’t seem especially concerned about this possibility. “I’ve come to the conclusion that because information constantly increases, there’s never going to be privacy,” one of the company’s early investors told the New York Times. “Laws have to determine what’s legal, but you can’t ban technology. Sure, that might lead to a dystopian future or something, but you can’t ban it.”³⁶

《纽约时报》的文章引发了一场针对该公司的争议风暴，也引起了黑客的注意，他们成功入侵 Clearview 的服务器，获得了该公司付费客户以及使用该应用程序 30 天免费试用版的潜在客户的完整名单。事实证明，Clearview 的用户包括联邦调查局、国际刑警组织、美国移民和海关执法局 (ICE) 和纽约南区美国检察官办公室等主要机构，以及全球数百个警察局。尽管该公司声称只与有资质的执法机构合作，但这款应用程序正在被百思买、梅西百货、Rite Aid 和沃尔玛等私营公司使用。更糟糕的是，有证据表明私营部门的员工可能在未经雇主授权的情况下使用该应用程序。BuzzFeed 的一项调查发现，与家得宝相关的五个账户使用该应用程序进行了近一百次搜索，尽管家得宝的管理层声称对此一无所知。³⁷换句话说，该技术的使用权已经渗透到更广泛的公共领域。

The Times article unleashed a storm of controversy focused on the company and also attracted the attention of hackers, who managed to break into Clearview’s servers and obtain a complete list of the company’s paying clients as well as prospective customers who were using a thirty-day free trial version of the app. It turned out that Clearview’s users included major agencies such as the FBI, Interpol, U.S. Immigration and Customs Enforcement (ICE) and the U.S. Attorney’s Office for the Southern District of New York, as well as hundreds of police departments across the globe. And despite the company’s claims that it worked only with credentialed law enforcement agencies, the app was being used at private companies including Best Buy, Macy’s, Rite Aid and Walmart. Worse still, there was evidence that private sector workers might be using the app without authorization from their employers. An investigation by BuzzFeed found that five accounts associated with Home Depot had performed nearly one hundred searches using the app, despite the fact that Home Depot’s management claimed to be completely unaware of this.³⁷ In other words, access to the technology was already seeping out into the broader public sphere.

曝光后立即引发了强烈反应。几周之内，Twitter、Facebook 和 Google 向该公司发出了停止令，要求其停止从其服务器抓取照片，并立即删除数据库中已有的任何图片。38到^二月底，苹果已禁用 Clearview 的 iPhone 应用程序，因为该公司绕过 App Store 违反了苹果的服务协议。39 此后不久^，该公司宣布将终止与私营公司的所有许可协议，专注于执法机构，但这一举措被广泛认为不够。五月，美国公民自由联盟对 Clearview 提起诉讼，宣称该公司的技术构成了“噩梦般的场景”，如果不加以阻止，“我们所熟知的隐私将终结”。40 ^Clearview 仍在运营，并表示它认为自己有权在互联网上搜索照片，并准备与社交媒体公司就此类访问权展开法律斗争。

The publicity brought about an immediate backlash. Within weeks, Twitter, Facebook and Google had sent cease-and-desist orders to the company demanding that it stop scraping photographs from their servers and immediately delete any images already in its database.³⁸ By the end of February, Apple had disabled Clearview’s iPhone app because the company had violated Apple’s service agreement by circumventing the App Store.³⁹ Shortly thereafter, the company announced it would terminate all licensing agreements with private companies and focus exclusively on enforcement agencies, but this was widely dismissed as insufficient. In May, the American Civil Liberties Union filed a lawsuit against Clearview and declared that the company’s technology posed “a nightmare scenario” that would “end privacy as we know it if it isn’t stopped.”⁴⁰ Clearview continues to operate and has stated that it believes it has the right to scour the internet for photographs and is prepared to do legal battle with the social media companies over such access.

Clearview AI 不仅为面部识别，而且为更广泛的人工智能提供了重要的警示。利用如此独特的强大技术，最小的技术专家团队——甚至可能是单个人——都可能引发几乎难以想象的社会或经济混乱，正如我们将在下一章中看到的那样，风险绝不仅限于部署人工智能监控技术。

Clearview AI offers an important cautionary tale not just for facial recognition but for artificial intelligence more generally. Wielding such a uniquely powerful technology, the smallest team of technical experts—or perhaps even a single individual—could conceivably unleash social or economic disruption on a nearly unimaginable scale, and as we’ll see in the next chapter, the risks are by no means confined to the deployment of AI-enabled surveillance technologies.

鉴于对该公司的巨大反对，Clearview 的野心似乎可能会受到遏制。但更普遍的是，西方国家正在迅速加速部署面部识别技术，民主社会将面临越来越迫切的需要，需要做出基于价值的权衡，并面对围绕使用该技术的道德问题。伦敦是西方监控最严密的城市，人均闭路电视摄像头数量超过北京。伦敦于 2020 年初开始部署^{41 个} 面部识别系统。伦敦大都会警察局表示，这些系统只会寻找“定制”监视名单上的人，该名单由严重或暴力犯罪通缉犯组成。不过，该系统也可用于搜寻失踪的儿童和成年人^。42

Given the overwhelming backlash against the company, Clearview’s ambitions seem likely to be reined in. More generally, however, the deployment of facial recognition is accelerating rapidly throughout Western countries, and democratic societies will face an increasingly urgent need to make value-based trade-offs and confront the ethical issues that surround the use of the technology. In London, which is already by far the most surveilled Western city, with more CCTV cameras per capita then Beijing,⁴¹ facial recognition systems began to be deployed in early 2020. The Metropolitan Police stated that the systems would seek only people on a “bespoke” watch list consisting of individuals wanted for serious or violent crimes. However, the system could also be used to search for missing children and adults.⁴²

在美国，大约四分之一的警察局可以使用面部识别技术。该系统还广泛用于机场，用于搜寻已知的恐怖分子或罪犯，并越来越多地用于在安全检查过程中验证身份。在大多数情况下，就像伦敦的系统一样，该技术仅用于识别特定监视名单上的个人。然而，我们正逐渐接近 Clearview 应用程序所预示的那种过度反乌托邦的可能性，几乎任何人都可以被识别。根据乔治城大学法学院隐私与技术中心 2016 年的分析，联邦调查局维护的照片数据库包含约 1.17 亿人的照片，约占美国成年人口的一半。43许多图像来自各州保存的驾照照片，其中包括所有持有州颁发的身份证的居民，而不仅仅是那些因^犯罪而被通缉或有犯罪记录的人。不用说，该系统并不要求个人同意将他们的照片纳入其中，也没有办法退出该系统。

In the United States, about a quarter of police departments have access to facial recognition technology. The systems are also widely employed at airports, where they search for known terrorists or criminals and are increasingly used to verify identity during the security screening process. In most cases, as with the system in London, the technology is used to identify only individuals on specific watch lists. However, we are gradually inching closer to the possibility of the kind of overreaching dystopia heralded by Clearview’s app, where virtually anyone can be identified. The FBI maintains photographic databases that, according to a 2016 analysis by the Georgetown University Law School’s Center on Privacy and Technology, include images for about 117 million people—roughly half the adult population of the United States.⁴³ Many of the images come from driver’s license photographs maintained by the states and include all residents with identification cards issued by the state, not just those who are wanted for crimes or have criminal records. There is, needless to say, no requirement that individuals give consent for their photographs to be included, and no way to opt out from the system.

尽管对隐私的潜在威胁非常真实，但重要的是要认识到，如果部署得当且符合道德规范，面部识别系统确实会带来明确的好处。许多危险的罪犯已经使用该技术被逮捕。虽然在 Clearview 的案例中，我认为隐私问题显然超过了任何优势，但该应用程序确实逮捕了危险的罪犯，并被证明在识别性侵犯者和儿童色情制品传播者方面特别有效。在公共场所部署面部识别系统同样可以带来降低犯罪率的真正好处。伦敦大都会警察局说得没错，“我们都希望在一个安全的城市生活和工作：公众理所当然地希望我们使用广泛可用的技术来阻止犯罪分子。” ⁴⁴

Though the potential threat to privacy is very real, it’s important to recognize that, properly and ethically deployed, facial recognition systems do bring unambiguous benefits. Many dangerous criminals have been apprehended using the technology. While in the case of Clearview, I would say the privacy issues pretty clearly outweigh any advantages, the app nonetheless did lead to the apprehension of dangerous criminals and proved to be particularly effective at identifying sexual predators and purveyors of child pornography. Facial recognition systems deployed in public places likewise can offer real benefits in terms of lower crime rates. London’s Metropolitan Police are not wrong when they say, “We all want to live and work in a city which is safe: the public rightly expect us to use widely available technology to stop criminals.”⁴⁴

事实上，尽管从西方的角度来看，中国正在广泛部署的监控显然是一种压迫，但大多数中国公众并不一定对此持负面看法。许多襄阳居民非常支持横穿马路的制度，因为它确实有效，曾经危险的十字路口现在变得更加有序。我曾亲自与许多居住在中国的人交谈过，他们反复提到的一个观察结果是，人们的安全感增强，远离犯罪，尤其是年幼孩子的父母更加安心。这一点的潜在重要性不应被低估。大多数人非常看重社区的安全感，这与更好的身心健康有关。在这方面，中国在许多情况下可以说比美国做得更好。

Indeed, the widespread surveillance being deployed in China, while clearly oppressive from a Western perspective, is not necessarily viewed in a negative way by most of the Chinese public. Many residents of Xiangyang are very supportive of the jaywalking system because it worked, and the once-dangerous intersection is now more orderly. I have personally talked to a number of people who live in China, and one observation that comes up again and again is an increased sense of safety from crime and, in particular, peace of mind for parents of young children. The potential importance of this should not be underestimated. A sense of security in one’s neighborhood is valued highly by most people and is correlated with better physical and mental health. This is one area where China, in many cases, is arguably outperforming the United States.

安全的环境对孩子来说尤其重要。纽约大学作家兼教授乔纳森·海特 (Jonathan Haidt) 一直大力提倡“自由放养”式的育儿方式。海特认为，在美国，我们创造了一种对孩子过度保护的文化，这种文化很可能剥夺了孩子获得无监督体验的重要机会，而这种体验有助于他们成长为自信的成年人。45^对于大多数美国父母来说，让年幼的孩子在无人看管的情况下步行上学或在社区公园玩耍是一个可怕的想法，在某些地方甚至是非法的。我怀疑，中国的年幼儿童并不特别了解这个过度干预的奥威尔式国家。然而，他们确实知道他们可以步行上学或在公园玩耍。如果中国压迫性的监控系统最终有一线希望，至少对最年轻的公民而言，那将是非常讽刺的。随着时间的推移，这可能实际上有助于培养出更具冒险精神和创新精神的年轻人一代。没有人希望美国采用中国的系统，但就基于人工智能的监控技术可以降低犯罪率并创造更安全的环境而言，应该仔细考虑其中的利弊。

A safe environment is especially important for children. Jonathan Haidt, an author and professor at New York University, has been a strong advocate for “free range” parenting. Haidt argues that, in the United States, we have created a culture that is becoming dangerously overprotective of children and likely robs them of important opportunities to have the kind of unsupervised experiences that will help them develop into confident adults.⁴⁵ For most American parents, the idea of allowing young children to walk to school or play in a neighborhood park without oversight is a terrifying, and in some places illegal, prospect. I suspect that young children in China are not especially aware of the overreaching Orwellian state. They do know, however, that they can walk to school or play in the park. It would be extraordinarily ironic if China’s oppressive system of surveillance turns out to have a silver lining, at least where its youngest citizens are concerned. In time, that might actually help produce a more adventurous and innovative generation of young people. No one wants China’s system in the United States, but to the extent that AI-based surveillance technologies can drive down crime rates and create safer environments, the trade-offs should be given careful consideration.

虽然人脸识别能给社会带来真正的好处，但公平地应用这项技术并确保其对不同人口群体的影响平等至关重要。而这里存在一个大问题。多项研究显示，人脸识别系统始终存在一定程度的种族和性别偏见。诚然，这与中国专门为寻找维吾尔族人而设计的算法不同，而是由于用于训练深度学习算法的数据集中白人男性面孔占多数。在一个常用的训练数据集中，白人占 83%，男性占 77%。46这个问题通常表现为非白人和女性面孔出现“假阳性”结果的可能性增加。换句话说^，女性和有色人种更有可能产生错误匹配。

Though facial recognition can bring real benefits to society, it’s critical that the technology be applied fairly and that the impact is equitable across demographic groups. And, here, there is a major problem. In a number of studies, facial recognition systems have consistently been shown to exhibit some degree of both racial and gender bias. To be sure, this is nothing like the Chinese algorithms explicitly designed to seek out Uyghurs, but rather results from a preponderance of white male faces in the datasets used to train deep learning algorithms. The faces in one commonly used training dataset are eighty-three percent white and seventy-seven percent male.⁴⁶ The problem generally manifests as an increased likelihood of a “false positive” result for non-white and female faces. In other words, women and people of color are more likely to generate an incorrect match.

2018 年，美国公民自由联盟将美国国会所有 538 名成员的照片与大量被捕人员入狱照片数据集进行了比较。美国公民自由联盟使用了 Amazon Web Services 提供的 Rekognition 系统，该系统在警察部门中越来越受欢迎，因为它的价格非常低。美国公民自由联盟仅花费约 12 美元就完成了这项实验。该系统将 28 名国会议员标记为被捕者，他们的照片包含在大头照数据集中。假设被捕者中没有一个人真正当选众议院或参议院议员，那么这些都是误报。除了错误数量众多之外，一个主要问题是系统生成的误报严重偏向非白人国会议员。有色人种约占国会议员的 20%，但他们占错误匹配的 39%。针对这项研究，亚马逊辩称，美国公民自由联盟对系统的配置不正确，因为它使用了默认的 80% 匹配置信度阈值，而不是更合适的 95%。然而，美国公民自由联盟指出，亚马逊没有提供关于正确设置的具体说明，许多警察部门可能会将系统设置为默认^设置。47

In 2018, the American Civil Liberties Union compared images of all 538 members of the U.S. Congress to a large dataset of booking photographs taken of people who had been arrested. The ACLU used the Rekognition system available through Amazon Web Services, which is becoming increasingly popular with police departments because it is available at a very low cost. The ACLU was able to conduct the experiment for only about $12. The system flagged twenty-eight members of Congress as arrestees whose photos were contained in the mug shot dataset. Assuming that none of the individuals who were arrested have, in fact, been elected to either the House or the Senate, these were all false positives. Aside from the sheer number of errors, a major concern was that the false positives generated by the system were weighted heavily toward non-white members of Congress. People of color represent about twenty percent of Congress, but they accounted for thirty-nine percent of the incorrect matches. In response to the study, Amazon argued that the ACLU had configured the system incorrectly because it used the default confidence threshold of eighty percent for matches, rather than a more appropriate ninety-five percent. The ACLU, however, noted that Amazon provides no specific instructions regarding the proper setting, and that many police departments would likely leave the system set to its defaults.⁴⁷

2019 年，美国商务部下属的国家标准与技术研究所 (NIST) 进行了一项更为全面的研究。NIST 评估了来自 99 家不同公司的 189 个面部识别系统。48^他们发现，在几乎所有情况下，欧洲人脸的误报率最低，而非洲人和亚洲人脸的误报率明显较高。可以预见的例外是中国公司开发的算法，东亚人脸的识别结果最为准确。这些系统对男性面孔的识别准确率通常高于女性面孔，尽管这种差异的幅度小于不同种族的识别准确率。

A much more comprehensive study was performed in 2019 by the National Institute of Standards and Technology (NIST), which is part of the U.S. Department of Commerce. NIST evaluated 189 facial recognition systems from ninety-nine different companies.⁴⁸ They found that, in nearly all cases, false positives were lowest for European faces and significantly higher for African and Asian faces. The predictable exception was for algorithms developed by Chinese companies, for which East Asian faces produced the most accurate results. The systems were also generally more accurate for male than for female faces, although the magnitude of the difference was smaller than for different races.

非白种人的准确率差异很大。例如，黑人的误报率可能比白人高出一百倍。换句话说，非裔美国人被错误地标记为潜在罪犯并因此受到不便、骚扰甚至拘留的可能性比白人高出一百倍。从本质上讲，这相当于非裔美国人已经熟悉的现实世界场景的数字化再现，例如，他们经常在零售店被保安跟踪或受到售货员的过度关注。

The difference in accuracy for non-white races was substantial. A black person, for example, might face a likelihood of a false positive more than a hundred times that of a white person. In other words, an African American might be a hundred times more likely than a white person to be incorrectly flagged as a potential offender and thereby inconvenienced, accosted or perhaps even detained. Essentially, this amounts to a digital rendition of the real-world scenarios already familiar to African Americans, where, for example, they are often followed by security staff in retail stores or given undue attention by sales clerks.

理论上，只需在训练数据集中加入更多不同的面孔，就可以解决这个问题。然而，开发面部识别系统的公司往往很难找到合乎道德且获得同意的高质量非白人面孔图像——换句话说，没有像 Clearview 那样使用从互联网上抓取图像等技术。49^这个问题的解决方案有时会引发问题，而在这个领域，愿意突破道德界限的公司有时可能会获得优势。2018 年，中国独角兽云从科技与津巴布韦政府达成了一项有争议的协议，为该国建立一个全面的面部识别系统。作为协议的一部分，云从科技将获得津巴布韦公民的照片，并将能够使用它们来训练其机器学习算法。由此产生的系统可能会部署在世界任何地方，当然，津巴布韦公民对此并不知情或不同意^。50

In theory, it should be possible to address this problem simply by including more diverse faces in the training datasets. However, the companies that develop facial recognition systems have often struggled to find high-quality images of non-white faces obtained ethically and with consent—or in other words, without resorting to techniques like scraping images from the internet, as was done by Clearview.⁴⁹ Solutions to this problem can sometimes raise questions of their own, and this is an area where companies willing to push ethical boundaries may sometimes obtain an advantage. In 2018, the Chinese unicorn CloudWalk entered into a controversial agreement with the government of Zimbabwe to build a comprehensive facial recognition system for the country. As part of the agreement, CloudWalk will obtain access to photographs of Zimbabwean citizens and will be able to use them to train its machine learning algorithms. The resulting systems could potentially be deployed anywhere in the world, without, of course, any knowledge or consent from the citizens of Zimbabwe.⁵⁰

这些问题以及 Clearview 的情况都清楚地表明，面部识别不能交给不受监管的私营部门。对该技术的监管和监督至关重要。如果《纽约时报》没有曝光该公司，Clearview 的技术可能早在公众意识到它代表的隐私威胁之前就已在没有任何监督的情况下传播到公众。至少，我们需要明确的法规来确保所部署算法的公平性，以及防止以威胁公众隐私的方式部署监控系统的保障措施。

Issues like these, as well as the situation with Clearview, make it clear that facial recognition cannot be left in the hands of an unregulated private sector. Regulation and oversight of the technology is essential. Had the New York Times not exposed the company, Clearview’s technology might have reached the public without any oversight whatsoever, long before there was any general awareness of the threat to privacy that it represented. At a minimum, we need clear regulations to ensure the fairness of any algorithms deployed, as well as safeguards that will prevent deployment of surveillance systems in ways that threaten public privacy.

由于缺乏通用标准，旧金山等一些司法管辖区已经主动全面禁止警察局和地方政府使用面部识别技术。然而，这并不包括私营部门。与中国一样，面部识别越来越普遍地被用作大型住宅开发项目的进入机制。一些居民提起诉讼，称这侵犯了他们的隐私。零售店同样可以不受任何限制地部署这项技术。显然，我们需要在国家层面制定法规，定义一套基本规则，这些规则必须适用于公共或私人部署的系统。对隐私、监控和公共安全重要性的态度各不相同，各个国家、地区和城市很可能会在面部识别和其他基于人工智能的监控技术的风险价值主张方面做出不同的权衡。在民主社会中，应该有一个透明的程序，纳入公众的意见，这项技术需要由一套保护所有相关人员权利的基本原则来管理。

In the absence of general standards, some jurisdictions, such as San Francisco, have already taken the initiative to completely ban the use of facial recognition technology by police departments and local government. However, this does not extend to the private sector. As in China, it’s becoming increasingly common for facial recognition to be deployed as an entry mechanism for large housing developments. Some residents have filed lawsuits arguing that this represents an invasion of their privacy. Retail stores likewise can deploy the technology with few constraints. Clearly, we need regulation at the national level that defines a basic set of rules that must apply to systems deployed either publicly or privately. Attitudes toward privacy, surveillance and the importance of public safety vary, and it seems very likely that individual countries, regions and cities will make different trade-offs around the risk-value proposition of facial recognition and other AI-based surveillance technologies. In democratic societies, there should be a transparent process that incorporates public input, and the technology needs to be governed by a set of foundational principles that protect the rights of all involved.

随着人工智能技术不断进步，与中国展开人工智能军备竞赛的可能性非常大，个人隐私面临前所未有的威胁，新的歧视形式也随之出现，而这只是其中的一部分。在下一章中，我们将更广泛地审视与人工智能相关的一些风险，并讨论哪些危险需要我们立即关注，哪些是可能在遥远的未来才会出现的更具推测性的担忧。

THE VERY REAL possibility of an AI arms race with China, unprecedented threats to personal privacy and new forms of discrimination are just a few of the emerging dangers as the technology of artificial intelligence continues to relentlessly advance. In the next chapter, we’ll take a broader view of some of the risks that are inherently coupled with AI and discuss which dangers need our immediate attention and which are more speculative concerns likely to arise only in the far future.

脚注

Footnote

ⁱ如果您对深度神经网络在语言翻译方面的应用能力存有疑虑，请比较一下中国《新一代人工智能发展规划》中的两个介绍部分。一个是谷歌对中国政府原始文件的机器翻译。另一个是由四名语言学家组成的团队进行专业翻译。

ⁱ If you have any doubts about the power of deep neural networks when applied to language translation, compare these two introductory sections from China’s “New Generation Artificial Intelligence Development Plan.” One is a Google machine translation of the original Chinese government document. The other was professionally translated by a team of four linguists.

以下是每份文件的第一段。你能分辨出哪个是哪个吗？

The first paragraph from each document is below. Can you tell which is which?

答：人工智能快速发展将深刻改变人类社会和世界。为抓住人工智能发展重大战略机遇，构建我国人工智能发展先发优势，加快建设创新型国家和世界科技强国，根据党中央、国务院的部署要求，制定本规划。

A. The rapid development of artificial intelligence will profoundly change human society and the world. In order to seize the major strategic opportunities in the development of artificial intelligence, build the first-mover advantage in the development of artificial intelligence in China, and accelerate the construction of an innovative country and a world power of science and technology, this plan was formulated in accordance with the deployment requirements of the Party Central Committee and the State Council.

(二)人工智能快速发展将深刻改变人类社会生活、改变世界。为抓住人工智能发展重大战略机遇，构建我国人工智能发展先发优势，加快建设创新型国家和世界科技强国，按照党中央、国务院要求，制定本规划。

B. The rapid development of artificial intelligence (AI) will profoundly change human society and life and change the world. To seize the major strategic opportunity for the development of AI, to build China’s first-mover advantage in the development of AI, to accelerate the construction of an innovative nation and global power in science and technology, in accordance with the requirements of the CCP Central Committee and the State Council, this plan has been formulated.

答案是 B 是人工翻译的版本。（见尾注 8。）

The answer is that B is the human-translated version. (See endnote 8.)

⁵⁰

第八章

CHAPTER 8 人工智能风险

AI RISKS

现在是十一月初，距离美国总统大选只有两天时间。这位民主党候选人在其职业生涯的大部分时间里都在为推进民权和扩大对边缘化社区的保护而奋斗。她在这个问题上的记录似乎无可挑剔。因此，当一段据称是候选人进行私人谈话的录音出现并立即在社交媒体上疯传时，人们感到非常震惊。在谈话中，候选人不仅使用了明显的种族主义语言，还公开承认，甚至嘲笑她一生都在隐藏偏见方面取得的成功。

IT’S EARLY NOVEMBER, JUST TWO DAYS BEFORE THE PRESIDENTIAL election in the United States. The Democratic candidate has spent much of her career fighting to advance civil rights and expand protections for marginalized communities. Her record on this issue is seemingly impeccable. It therefore comes as an immense shock when an audio recording that purports to be of the candidate engaging in a private conversation appears, and then immediately goes viral, on social media. In the conversation, the candidate not only uses explicitly racist language, she also openly admits, and even laughs about, her lifelong success in keeping her bigotry hidden.

音频片段出现后不到一小时，候选人就极力否认其真实性。认识她的人都不相信这些话可能是她说的，有数十人站出来支持她。然而，任何选择相信她的人都必须面对一个非常令人不安的现实： 这是她的声音。或者至少在任何人听来，这听起来都是候选人在说话。她发音的独特方式、她讲话的节奏，都似乎无可否认地属于大多数人期待很快成为美国当选总统的那位女性。

Within an hour of the audio clip’s appearance, the candidate vehemently denies its authenticity. No one who knows her personally believes that the words can possibly be hers, and dozens of people come forward to support her. Anyone who chooses to believe her, however, has to confront a very uncomfortable reality: it is her voice. Or at least to virtually any human ear, it appears to be the candidate speaking. The distinct way in which she enunciates certain words and phrases, the cadence of her speech, all seem to undeniably belong to the woman that most people expect will soon become the president-elect of the United States.

随着这段录音在互联网上爆红，并在有线电视上反复播放，社交媒体世界陷入混乱和愤怒。在获得提名之前，这位候选人进行了一场激烈的初选之战，现在她的对手的一些愤怒支持者开始要求她下台。

As the audio recording explodes on the internet and is played repeatedly on cable TV, the social media universe reels with confusion and outrage. Before securing the nomination, the candidate fought a vicious primary battle, and now some angry supporters of her opponents begin to call for her to step aside.

竞选团队立即聘请了一组专家对音频文件进行独立审查。经过一天的深入分析，他们宣布这段录音很可能是“深度伪造”——由机器学习算法生成的音频，这些算法已经对候选人的讲话样本进行了广泛的训练。多年来，一直有人警告人们不要使用深度伪造，但到目前为止，它们还很初级，很容易被识别为伪造。这个例子不同；很明显，这项技术已经取得了显著的进步。即使是专家小组也不能绝对肯定地说这段音频文件是伪造的，而不是真实的录音。

The campaign immediately hires a panel of experts to independently review the audio file. After a day of intense analysis, they declare that the recording is likely a “deepfake”—audio generated by machine learning algorithms that have been extensively trained on examples of the candidate speaking. There have been warnings about deepfakes for years, but so far they have been rudimentary and easy to identify as fabrications. This example is different; it is clear that the state of the technology has advanced significantly. Even the panel of experts cannot state with absolute certainty that the audio file is a fake and not an actual recording.

在专家组的判断下，竞选团队成功删除了大部分在线音频文件。然而，数百万人已经听到了这些话。随着选举日的临近，一些关键问题浮出水面：所有听过录音的人是否都知道这可能是假的？被告知录音是伪造的选民能否设法以某种方式“忘记”那些已经不可磨灭地刻在他们记忆中的仇恨言论——尤其是如果他们恰好属于对话中针对的群体？这段录音会降低民主党候选人最依赖的社区的投票率吗？如果她输了，大多数美国人会觉得选举被窃取了吗？到那时会发生什么？

Based on the determination of the expert panel, the campaign succeeds in having most online copies of the audio file taken down. However, millions have already heard the words. As Election Day dawns, a number of critical questions loom: Has everyone who heard the recording learned that it is likely a fake? Can voters who are told the recording is a fabrication manage to somehow “unhear” the hateful words that are by now indelibly etched into their memories—especially if they happen to belong to a group targeted in the conversation? Will the audio clip depress turnout within the communities that the Democratic candidate most relies on? And if she loses, will a majority of the American people feel that the election has been stolen? What will happen then?

虽然上述场景显然是虚构的，但现实是，类似我所描述的事情可能会发生——也许在短短几年内。如果你对此表示怀疑，请考虑一下，2019 年 7 月，网络安全公司赛门铁克透露，三家未具名的公司已被犯罪分子利用音频深度伪造技术骗取数百万美元。1^在这三起案件中，犯罪分子都是使用人工智能生成的公司首席执行官的音频片段来伪造电话，命令财务人员将钱转入非法银行账户。首席执行官——就像上面想象的总统候选人一样——通常拥有丰富的在线音频数据（演讲、电视露面等），可用于训练机器学习算法。由于技术尚未发展到可以产生真正高质量音频的程度，因此在这些案件中，犯罪分子故意插入背景噪音（例如交通噪音）来掩盖瑕疵。然而，未来几年深度伪造的质量肯定会大幅提高，最终可能会达到真相与虚构几乎难以区分的地步。

While the scenario above is obviously fiction, the reality is that something similar to what I have described could happen—perhaps within just a few years. If you doubt this, consider that, in July 2019, the cybersecurity firm Symantec revealed that three unnamed corporations had already been bilked out of millions of dollars by criminals using audio deepfakes.¹ In all three cases, the criminals did this by using an AI-generated audio clip of the company CEO’s voice to fabricate a phone call ordering financial staff to move money to an illicit bank account. CEOs—like the presidential candidate imagined above—typically have a rich trove of online audio data (speeches, television appearances, etc.) that could be used to train machine learning algorithms. Because the technology is not yet at the point where it can produce truly high-quality audio, the criminals in these cases intentionally inserted background noise (such as traffic) to mask the imperfections. However, the quality of deepfakes is certain to get dramatically better in the coming years, and eventually, things will likely reach a point where truth is virtually indistinguishable from fiction.

恶意部署深度伪造技术只是我们在人工智能发展过程中面临的重要风险之一，这种技术不仅可用于生成音频，还可生成照片、视频甚至连贯的文本。在上一章中，我们看到了人工智能监控和面部识别技术如何摧毁个人隐私的概念，并带领我们走向奥威尔式的未来。在本章中，我们将探讨随着人工智能变得越来越强大，可能出现的其他一些主要问题。

Malicious deployment of deepfakes, which can be used to generate not only audio but also photographs, video and even coherent text, is just one of the important risks we face as artificial intelligence advances. In the previous chapter, we saw how AI-enabled surveillance and facial recognition technologies could destroy the very concept of personal privacy and lead us into an Orwellian future. In this chapter, we will look at some of the other major concerns that are likely to arise as AI becomes ever more powerful.

什么是真实，什么是幻觉？深度伪造和安全威胁

WHAT IS REAL, AND WHAT IS ILLUSION? DEEPFAKES AND THREATS TO SECURITY

深度伪造通常由深度学习的一项创新技术“生成对抗网络”（GAN）驱动。GAN 在一种游戏中部署了两个相互竞争的神经网络，这种游戏不断推动系统生成更高质量的模拟媒体。例如，用于生成假照片的 GAN 将包括两个集成的深度神经网络。第一个网络称为“生成器”，用于生成伪造的图像。第二个网络在由真实照片组成的数据集上进行训练，称为“鉴别器”。生成器合成的图像与真实照片混合后输入鉴别器。这两个网络不断相互作用，进行竞赛，其中鉴别器评估生成器生成的每张照片并判断其是真是假。生成器的目标是通过伪造照片来欺骗鉴别器。随着两个网络不断进行迭代战斗，图像质量越来越好，直到系统最终达到一种平衡，其中鉴别器只能猜测它所分析图像的真实性。这种技术可以制作出令人惊叹的伪造图像。在互联网上搜索“GAN 假脸”，你会发现许多高分辨率图像的例子，这些图像描绘了完全不存在的人。试着扮演鉴别器网络的角色。这些照片看起来完全真实，但它们是一种幻觉——一种从数字以太中变幻出来的渲染。

Deepfakes are often powered by an innovation in deep learning known as a “generative adversarial network,” or GAN. GANs deploy two competing neural networks in a kind of game that relentlessly drives the system to produce ever higher quality simulated media. For example, a GAN designed to produce fake photographs would include two integrated deep neural networks. The first network, called the “generator,” produces fabricated images. The second network, which is trained on a dataset consisting of real photographs, is called the “discriminator.” The images synthesized by the generator are mixed with actual photographs and fed to the discriminator. The two networks interact continuously, engaging in a contest in which the discriminator evaluates each photograph produced by the generator and decides whether it is real or fake. The generator’s objective is to try to deceive the discriminator by passing off faux photographs. As the two networks continue their iterative battle, the image quality gets better and better until ultimately the system reaches a kind of equilibrium in which the discriminator can do no better than guess at the authenticity of the images it analyzes. This technique produces astonishingly impressive fabricated images. Search the internet for “GAN fake faces,” and you’ll find numerous examples of high-resolution images that portray completely nonexistent individuals. Try stepping into the role of the discriminator network. The photographs seem completely real, but they are an illusion—a rendering conjured up from the digital ether.

生成对抗网络是由蒙特利尔大学研究生 Ian Goodfellow 发明的。2014 年的一个晚上，Goodfellow 和几个朋友去了当地的一家酒吧。他们讨论了如何构建一个可以生成高质量图像的深度学习系统。在喝了不知多少杯啤酒后，Goodfellow 提出了生成对抗网络的基本概念，但遭到了极大的质疑。之后，Goodfellow 回家后立即开始编码。几个小时内，他就制作出了第一个可以运行的 GAN。这一成就让 Goodfellow 成为深度学习社区的传奇人物。Facebook 首席人工智能科学家 Yann LeCun 表示，生成对抗网络是“过去 20 年深度学习领域最酷的想法”。2^在蒙特利尔大学获得博士学位后，Goodfellow 继续在 Google Brain 项目和 OpenAI 工作，现在是 Apple 的机器学习总监。他也是深度学习领域顶尖大学教科书的主要作者。

Generative adversarial networks were invented by a graduate student at the University of Montreal named Ian Goodfellow. One evening in 2014, Goodfellow went out to a local bar with a few of his friends. They discussed the problem of building a deep learning system that could generate high-quality images. After imbibing an unknown number of beers, Goodfellow proposed the basic concept behind a generative adversarial network, but was met with extreme skepticism. Afterward, Goodfellow went home and immediately began coding. Within hours, he had the first functioning GAN. The accomplishment would turn Goodfellow into a legendary figure within the deep learning community. Yann LeCun, the chief AI scientist at Facebook, says generative adversarial networks are “the coolest idea in deep learning in the last 20 years.”² After completing his PhD at the University of Montreal, Goodfellow went on to work at the Google Brain project and OpenAI and is now a director of machine learning at Apple. He is also the primary author of the leading university textbook on deep learning.

生成对抗网络可以以多种积极的方式部署。特别是，合成图像或其他媒体可以用作其他机器学习系统的训练数据。例如，使用 GAN 创建的图像可用于训练自动驾驶汽车中使用的深度神经网络。当无法以合乎道德的方式获得足够数量的有色人种真实高质量图像时，也有人提议使用合成的非白人面孔来训练面部识别系统，以此作为克服种族偏见问题的一种方式。当应用于语音合成时，GAN 可用于为失去说话能力的人提供听起来像自己声音的计算机生成的替代声音。已故的斯蒂芬霍金因神经退行性疾病 ALS 或卢格里格氏症而失去声音，他以独特的计算机合成声音说话而闻名。最近，像 NFL 球员蒂姆肖这样的 ALS 患者通过对患病前的录音进行深度学习系统训练，恢复了他们的自然声音。

Generative adversarial networks can be deployed in many positive ways. In particular, synthesized images or other media can be used as training data for other machine learning systems. For example, images created with a GAN might be used to train the deep neural networks used in self-driving cars. There have also been proposals to use synthetic non-white faces to train facial recognition systems as a way of overcoming the racial bias problem when sufficient numbers of high-quality images of real people of color cannot be ethically obtained. When applied to voice synthesis, GANs can be used to provide people who have lost the ability to speak with a computer-generated replacement that sounds like their own voice. The late Stephen Hawking, who lost his voice to the neurodegenerative disease ALS, or Lou Gehrig’s disease, famously spoke in a distinctive computer synthesized voice. More recently, ALS patients like the NFL player Tim Shaw have had their natural voices restored by training deep learning systems on recordings made before the illness struck.

然而，该技术被恶意使用的可能性是不可避免的，而且有证据表明，对于许多精通技术的人来说，这是无法抗拒的。为幽默或教育目的而制作的广泛流传的深度伪造视频展示了什么是可能的。你可以找到许多假视频，其中有马克·扎克伯格等知名人士说了一些他们可能永远不会说的话——至少在公开场合不会说。最著名的例子之一是由演员兼喜剧演员乔丹·皮尔与 BuzzFeed合作制作的，他因模仿巴拉克·奥巴马的声音而闻名。在皮尔的公益视频中，奥巴马旨在让公众意识到深度伪造带来的迫在眉睫的威胁，视频中奥巴马说了这样的话：“特朗普总统是个彻头彻尾的白痴。” ³在这种情况下，声音是皮尔模仿奥巴马的，使用的技术通过操纵奥巴马总统的嘴唇来改变现有视频，使其与皮尔的讲话同步。最终，我们可能会看到这样的视频，其中的声音也是深度伪造的。

However, the potential for malicious use of the technology is inescapable and, evidence already suggests for many tech savvy individuals, irresistible. Widely available deepfake videos created with humorous or educational intent demonstrate what is possible. You can find numerous fake videos featuring high-profile individuals like Mark Zuckerberg saying things they would presumably never say—at least in public. One of the most famous examples was created by the actor and comedian Jordan Peele, who is known for his impersonation of Barak Obama’s voice, in collaboration with BuzzFeed. In Peele’s public service video intended to make the public aware of the looming threat from deepfakes, Obama says things like “President Trump is a total and complete dipshit.”³ In this instance, the voice is Peele’s imitation of Obama, and the technique used alters an existing video by manipulating President Obama’s lips so they synchronize with Peele’s speech. Eventually, we will likely see videos like this in which the voice is also a deepfake fabrication.

一种尤为常见的深度伪造技术能够将一个人的脸部数字转移到另一个人的真实视频中。据提供深度伪造检测工具的初创公司 Sensity（前身为 Deeptrace）称，2019 年网上发布了至少 15,000 件深度伪造作品，比上一年增长了 84%。4^其中，整整 96% 涉及色情图片或视频，其中将名人（几乎总是女性）的脸移植到色情演员的身上。5虽然泰勒·斯威夫特和斯嘉丽·约翰逊等名人一直是主要目标，^但这种数字滥用最终可能针对几乎任何人，尤其是随着技术的进步以及制作深度伪造的工具变得越来越普及和易于使用。

An especially common deepfake technique enables the digital transfer of one person’s face to a real video of another person. According to the startup company Sensity (formerly Deeptrace), which offers tools for detecting deepfakes, there were at least 15,000 deepfake fabrications posted online in 2019, and this represented an eighty-four percent increase over the prior year.⁴ Of these, a full ninety-six percent involve pornographic images or videos in which the face of a celebrity—nearly always a woman—is transplanted onto the body of a pornographic actor.⁵ While celebrities like Taylor Swift and Scarlett Johansson have been the primary targets, this kind of digital abuse could eventually be targeted against virtually anyone, especially as the technology advances and the tools for making deepfakes become more available and easier to use.

随着深度伪造的质量不断提高，伪造的音频或视频媒体可能真正造成破坏，这似乎是一种不可避免的威胁。正如本章开头的虚构轶事所示，一个足够可信的深度伪造可能会彻底改变历史的轨迹——而制造这种伪造的手段可能很快就会落入政治人物、外国政府甚至是淘气的青少年手中。而且，需要担心的不仅仅是政客和名人。在病毒式视频、社交媒体羞辱和“抵制文化”的时代，几乎任何人都可能成为目标，并可能因深度伪造而毁掉自己的职业生涯和生活。由于种族不平等的历史，美国可能特别容易受到精心策划的社会和政治动荡的影响。我们已经看到，描述警察暴力的病毒式视频几乎可以立即引发大规模抗议和社会动荡。绝非不可想象，在未来的某个时候，一段极具煽动性、足以破坏社会结构的视频可能会被合成——也许是由外国情报机构制作。

As the quality of deepfakes relentlessly advances, the potential for fabricated audio or video media to be genuinely disruptive looms as a seemingly inevitable threat. As the fictional anecdote at the beginning of this chapter illustrates, a sufficiently credible deepfake could quite literally shift the arc of history—and the means to create such fabrications might soon be in the hands of political operatives, foreign governments or even mischievous teenagers. And it’s not just politicians and celebrities who need to worry. In the age of viral videos, social media shaming and “cancel culture,” virtually anyone could be targeted and possibly have both their career and life destroyed by a deepfake. Because of its history of racial injustice, the United States may be especially vulnerable to orchestrated social and political disruption. We’ve seen how viral videos depicting police brutality can almost instantly lead to widespread protests and social unrest. It is by no means inconceivable that, at some point in the future, a video so inflammatory that it threatens to rend the very social fabric could be synthesized—perhaps by a foreign intelligence agency.

除了用于攻击或破坏的视频或声音片段之外，对于那些只想牟利的人来说，还有几乎无穷无尽的非法机会。犯罪分子将渴望利用这项技术进行各种活动，从金融和保险欺诈到股市操纵。一段公司首席执行官发表虚假声明或行为古怪的视频可能会导致该公司股价暴跌。深度伪造还会给司法系统带来麻烦。伪造的媒体可能会被当作证据，法官和陪审团最终可能会生活在一个很难甚至不可能知道他们亲眼所见是否真实的世界里。

Beyond videos or sound clips intended to attack or disrupt, there will be nearly endless illicit opportunities for those who simply want to profit. Criminals will be eager to employ the technology for everything from financial and insurance fraud to stock market manipulation. A video of a corporate CEO making a false statement, or perhaps engaging in erratic behavior, would likely cause the company’s stock to plunge. Deepfakes will also throw a wrench into the legal system. Fabricated media could be entered as evidence, and judges and juries may eventually live in a world where it is difficult, or perhaps impossible, to know whether what they see before their eyes is really true.

可以肯定的是，有聪明的人正在研究解决方案。例如，Sensity 销售的软件声称可以检测到大多数深度伪造。然而，随着技术的进步，不可避免地会出现一场军备竞赛——与制造新计算机病毒的人和销售软件来防范病毒的公司之间的竞赛没有什么不同——在这场竞赛中，恶意行为者可能总是至少占有一点优势。Ian Goodfellow 说，他认为我们无法仅通过“看像素”就能知道图像是真是假。6^相反，我们最终将不得不依靠照片和视频的控制论签名等身份验证机制。也许有一天，每台相机和手机都会在其记录的每段媒体中注入数字签名。一家名为 Truepic 的初创公司已经提供了一款应用程序来提供这种功能。该公司的客户包括大型保险公司，这些公司依靠客户发送的照片来记录从建筑物到珠宝和昂贵饰品等所有物品的价值。⁷ 不过，Goodfellow 认为，最终可能没有万无一失的技术解决方案来解决深度伪造问题。相反，我们必须以某种方式学会在一个前所未有的新现实中导航，在这个现实中，我们看到和听到的东西可能永远是一种幻觉。

To be sure, there are smart people working on solutions. Sensity, for example, markets software that it claims can detect most deepfakes. However, as the technology advances, there will inevitably be an arms race—not unlike the one between those who create new computer viruses and the companies that sell software to protect against them—in which malicious actors will likely always have at least a small advantage. Ian Goodfellow says he doesn’t think we will be able to know if an image is real or fake simply by “looking at the pixels.”⁶ Instead, we’ll eventually have to rely on authentication mechanisms like cybernetic signatures for photos and videos. Perhaps someday every camera and mobile phone will inject a digital signature into every piece of media it records. One startup company, Truepic, already offers an app to provide this kind of capability. The company’s customers include major insurance companies that rely on photographs that their customers send in to document the value of everything from buildings to jewelry and expensive trinkets.⁷ Still, Goodfellow thinks that ultimately there’s probably not going to be a foolproof technological solution to the deepfake problem. Instead, we will have to somehow learn to navigate within a new and unprecedented reality where what we see and what we hear can always potentially be an illusion.

虽然深度伪造的目的是欺骗人类，但相关问题涉及恶意伪造数据，旨在欺骗或控制机器学习算法。在这些“对抗性攻击”中，专门设计的输入会导致机器学习系统出错，从而使攻击者能够产生所需的输出。在机器视觉的情况下，这涉及在视野中放置某些东西，从而扭曲神经网络对图像的解释。在一个著名的例子中，研究人员拍摄了一张熊猫的照片，深度学习系统以大约 58% 的置信度正确识别了这只熊猫，并通过在图像中添加精心构建的视觉噪声，欺骗系统超过 99% 地确定熊猫是长臂猿。8一个特别令人不寒而栗的演示发现，只需在停车标志上添加四个小矩形黑白贴纸，就^可以欺骗自动驾驶汽车中使用的图像识别系统，使其相信停车标志实际上是 45 英里/小时的限速标志。⁹ 换句话说，对抗性攻击很容易造成生死攸关的后果。在这两种情况下，人类观察者甚至可能没有注意到偷偷添加到图像中的信息，当然也不会感到困惑。我认为这特别生动地证明了当今深度神经网络中凝聚的理解是多么肤浅和脆弱。

While deepfakes are intended to deceive human beings, a related problem involves the malicious fabrication of data intended to trick or gain control of machine learning algorithms. In these “adversarial attacks,” specially designed inputs cause a machine learning system to make an error in a way that allows the attacker to produce a desired output. In the case of machine vision, this involves placing something in the visual field that distorts the neural network’s interpretation of the image. In one famous example, researchers took a photograph of a panda, which a deep learning system identified correctly with about a fifty-eight percent confidence level, and by adding carefully constructed visual noise to the image, they tricked the system into being more than ninety-nine percent certain that the panda was instead a gibbon.⁸ An especially chilling demonstration found that simply adding four small rectangular black and white stickers to a stop sign tricked an image recognition system of the type used in self-driving cars into believing the stop sign was instead a 45 mph speed limit sign.⁹ In other words, an adversarial attack could easily have life-or-death consequences. In both of these cases, a human observer might not even notice—and certainly wouldn’t be confused by—the information surreptitiously added to the image. This is, I think, an especially vivid demonstration of just how shallow and brittle the understanding that coalesces in today’s deep neural networks really is.

对抗性攻击在人工智能研究界受到高度重视，并被视为一个关键漏洞。事实上，Ian Goodfellow 的大部分研究生涯都致力于研究机器学习系统中的安全问题并开发潜在的保护措施。构建能够抵御对抗性攻击的人工智能系统并非易事。一种方法涉及所谓的“对抗性学习”，换句话说，故意在训练数据中包含对抗性示例，希望神经网络能够在系统部署后识别出发生的攻击。然而，与深度伪造一样，这很可能注定是一场永无止境的军备竞赛，攻击者将永远占据优势。正如 Goodfellow 指出的那样，“还没有人设计出一种真正强大的防御算法，可以抵御各种对抗性示例攻击算法。” ¹⁰

Adversarial attacks are taken seriously within the AI research community and are viewed as a critical vulnerability. Indeed, Ian Goodfellow has devoted much of his research career to studying security issues within machine learning systems and developing potential safeguards. Building AI systems that are robust in the face of adversarial attacks is no easy task. One approach involves what is called “adversarial learning,” or in other words intentionally including adversarial examples in the training data in the hope that the neural network will be able to identify attacks if they occur once the system is deployed. As with deepfakes, however, there is likely destined to be a perpetual arms race in which attackers will always have an advantage. As Goodfellow points out, “no one has yet designed a truly powerful defense algorithm that can resist a wide variety of adversarial example attack algorithms.”¹⁰

对抗性攻击是机器学习系统特有的，但它们将成为网络犯罪分子、黑客或外国情报机构可以利用的计算机漏洞清单中更重要的一项。随着人工智能的广泛应用，以及物联网导致设备、机器和基础设施之间的互联互通日益加深，安全问题将变得更加严重，网络攻击几乎肯定会变得更加频繁。人工智能的广泛应用必然会导致系统更加自主，参与其中的人更少，而这些系统将成为越来越有吸引力的网络攻击目标。例如，想象一下，有一天自动驾驶卡车运送食品、药品和关键物资。一次让这些车辆停下来甚至造成严重延误的攻击很容易造成危及生命的后果。

Adversarial attacks are specific to machine learning systems, but they will become one more important item on the list of computer vulnerabilities that can be exploited by cybercriminals, hackers or foreign intelligence agencies. As artificial intelligence is increasingly deployed, and as the Internet of Things results in ever more interconnection between devices, machines and infrastructure, security issues will become vastly more consequential, and cyberattacks will almost certainly become more frequent. Wider deployment of AI will inevitably result in systems that are more autonomous, with fewer humans in the loop, and these systems will make increasingly attractive targets for cyberattack. Imagine, for example, that someday self-driving trucks deliver food, medicine and critical supplies. An attack that managed to bring these vehicles to a halt, or even create significant delays, could easily have life-threatening consequences.

所有这些的结果是，人工智能的可用性和依赖性增加将伴随着系统性安全风险，这将包括对关键基础设施和系统以及社会秩序、经济和民主制度的威胁。我认为，安全风险是人工智能崛起带来的最重要的近期危险。因此，我们必须投资于专注于构建强大的人工智能系统的研究，并在政府和商业部门之间建立有效的联盟，在出现严重漏洞之前制定适当的法规和保障措施。

The upshot of all this is that increased availability and reliance on artificial intelligence will come coupled with systemic security risk, and this will include threats to critical infrastructure and systems, as well as to the social order, our economy and our democratic institutions. Security risks are, I would argue, the single most important near-term danger associated with the rise of artificial intelligence. For this reason, it is critical that we invest in research focused on building AI systems that are robust, and that we form an effective coalition between government and the commercial sector to develop appropriate regulations and safeguards before critical vulnerabilities are introduced.

致命自主武器

LETHAL AUTONOMOUS WEAPONS

数百架微型无人机齐聚美国国会大厦，展开协同攻击。无人机利用面部识别技术识别特定人员，然后高速直冲他们，通过投放小型爆炸物进行有针对性的神风特攻队式暗杀，这种爆炸物的杀伤力与子弹一样有效。国会大厦一片混乱，但后来发现，所有被针对的国会议员都属于同一个政党。

Hundreds of miniature drones swarm through the U.S. Capitol building in a coordinated attack. Employing facial recognition technology, the drones identify specific individuals and then fly directly at them at high speed, carrying out targeted kamikaze assassinations by delivering a small shaped explosive that kills just as effectively as a bullet. The capitol is in complete chaos, but it later turns out that all the targeted members of Congress belonged to a single political party.

这只是 2017 年短片 《屠杀机器人^》中勾勒出的令人毛骨悚然的场景之一。11该视频旨在警告人们致命自主武器所带来的迫在眉睫的危险，由一个团队与加州大学伯克利分校计算机科学教授斯图尔特·拉塞尔合作制作，斯图尔特近期主要研究人工智能技术不断发展过程中固有的风险。拉塞尔认为，致命自主武器（联合国将其定义为能够“无需人工干预”定位、选择和消灭人类目标”的武器）¹²应归类为一种新型大规模杀伤性武器。换句话说，这些人工智能武器系统最终可能会像化学武器、生物武器甚至核武器一样具有破坏性和破坏性。

This is just one chilling scenario sketched out in the 2017 short film Slaughterbots.¹¹ Designed as a warning about the looming peril of lethal autonomous weapons, the video was produced by a team working with Stuart Russell, a professor of computer science at the University of California, Berkeley, who has focused much of his recent work on the inherent risks of artificial intelligence as the technology continues to advance. Russell believes that lethal autonomous weapons, which the United Nations defines as weapons that can “locate, select and eliminate human targets without human intervention,”¹² should be classified as a new type of weapon of mass destruction. In other words, these AI-powered weapons systems could ultimately be as disruptive and destabilizing as chemical, biological or perhaps even nuclear arms.

这一论点主要基于这样一个事实：一旦消除人类的直接控制和杀戮授权，这些武器在它们能够造成的破坏方面就变得高度可扩展。任何无人机都可能被用作武器，你可以一次发射数百架，但如果它们是远程控制的，你还需要数百人来驾驶这些设备。然而，如果无人机是完全自主的，一个小团队就可以部署大规模的无人机群，造成几乎难以想象的大屠杀。正如拉塞尔告诉我的那样，“有人可以发动一次袭击，控制室里的五个人就可以发射 10,000,000 件武器，消灭某个国家 12 岁到 60 岁之间的所有男性。所以，这些可能是大规模杀伤性武器，而且它们具有可扩展性。” ¹³ 鉴于面部识别算法能够根据种族、性别或服饰进行歧视，我们很容易想象出真正令人不寒而栗的场景，包括自动种族清洗或对政治对手的大规模暗杀，其残酷程度和速度是曾经难以想象的。

This argument is based primarily on the fact that once you eliminate direct human control and authorization to kill, such weapons become highly scalable in terms of the destruction they can unleash. Any drone could potentially be used as a weapon, and you could launch hundreds of them at a time, but if they are controlled remotely you would also need hundreds of people to pilot the devices. If the drones are fully autonomous, however, a small team could deploy massive swarms, unleashing almost unimaginable carnage. As Russell told me, “Someone can launch an attack, where five guys in a control room could launch 10,000,000 weapons and wipe out all males between the age of 12 and 60 in some country. So, these can be weapons of mass destruction, and they have this property of scalability.”¹³ Given the ability for facial recognition algorithms to discriminate on the basis of ethnicity, gender or attire, it’s easy to imagine truly chilling scenarios involving automated ethnic cleansing, or mass assassination of political opponents, carried out with a ruthlessness and speed that would have once been inconceivable.

即使我们可以完全抛开真正反乌托邦的可能性，假设这项技术严格限于合法的军事行动，自主武器也引发了严重的伦理问题。赋予机器独立夺取人类生命的能力在道德上是否可以接受，即使这样做可能会提高瞄准效率，并可能降低对无辜旁观者造成附带伤害的风险？在缺乏直接人类控制的情况下，如果发生导致伤亡的错误，谁应该承担责任？

Even if we could completely set aside the truly dystopian possibilities and assume the technology would be limited strictly to legitimate military engagements, autonomous weapons raise critical ethical concerns. Is it morally acceptable to give a machine the ability to independently take a human life, even if doing so might increase targeting efficiency and perhaps reduce the risk of collateral damage to innocent bystanders? And in the absence of direct human control, who should be held accountable in the event of an error that results in injury or loss of life?

他们正在推进的技术可能会被用在此类武器中，这一危险激起了众多人工智能研究人员的极大热情。超过 4,500 名个人以及数百家公司、组织和大学签署了公开信，宣布他们打算永远不研究自主武器，并呼吁全面禁止该技术。联合国常规武器公约正在发起一项倡议，以禁止完全自主的杀人机器，就像已经禁止化学武器和生物武器一样。然而，进展并不令人满意。根据致力于联合国禁令的倡导组织“制止杀人机器人运动”的数据，截至 2019 年，已有 29 个主要较小或发展中国家正式呼吁全面禁止自主武器技术。然而，主要军事大国并未加入。唯一的例外是中国，它签署了协议，但规定它只想禁止实际使用武器，允许开发和生产它们。¹⁴ 美国和俄罗斯都反对禁令，因此这些武器似乎不太可能在短期内被彻底禁止^。15

The danger that the technology they’re working to advance might be deployed in such weapons stirs a great deal of passion among many artificial intelligence researchers. More than 4,500 individuals as well as hundreds of companies, organizations and universities have signed open letters declaring their intention to never work on autonomous weapons and calling for a general ban on the technology. An initiative is underway within the United Nations Convention on Conventional Weapons to outlaw fully autonomous killing machines in much the same way that chemical and biological weapons are already forbidden. Progress, however, has been underwhelming. According to the Campaign to Stop Killer Robots, an advocacy group focused on a UN ban, as of 2019, twenty-nine mostly smaller or developing nations have formally called for a complete prohibition of autonomous weapons technology. The major military powers, however, are not on board. The one exception is China, which has signed on with the stipulation that it wants to ban only actual use of the weapons, allowing for their development and production.¹⁴ The United States and Russia have both opposed a ban, and it therefore seems unlikely that the weapons will be completely outlawed anytime soon.¹⁵

我个人的看法比较悲观。在我看来，大国之间的竞争态势和缺乏信任至少可能会使全自动武器的发展成为几乎确定的事情。事实上，美国军方的每个部门以及包括俄罗斯、中国、英国和韩国在内的国家都在积极开发具有集群能力的无人机。16同样，美国陆军正在部署看起来像小型坦克的武装机器人，^{17据报道，空军正在开发一种无人驾驶的人工智能战斗机，能够在空中格斗中击败}^人类驾驶的飞机。18^中国、俄罗斯、以色列和其他国家也在部署或开发类似的技术^。19

My own view is rather pessimistic. It seems to me that the competitive dynamic and lack of trust between major countries will probably make at least the development of fully autonomous weapons a near certainty. Indeed, every branch of the U.S. military as well as nations including Russia, China, the United Kingdom and South Korea are actively developing drones with the ability to swarm.¹⁶ Likewise, the U.S. Army is deploying armed robots that look like small tanks,¹⁷ and the Air Force is reportedly developing an unmanned AI-driven fighter jet with the ability to defeat human-piloted aircraft in dogfights.¹⁸ China, Russia, Israel and other countries are also deploying or developing similar technologies.¹⁹

到目前为止，美国和其他主要军队已承诺将始终让人类了解情况，在这些机器发动可能导致生命损失的攻击之前，必须获得具体授权。然而，现实是战场上的全面自动化将带来巨大的战术优势。没有人能够以与人工智能相当的速度做出反应和决策。一旦一个国家违反目前对完全自主的非正式禁令并开始部署这些能力，任何竞争对手的军队都不可避免地必须立即效仿，否则就会处于严重劣势。这种对落后的担忧可能是美国、中国和俄罗斯都反对正式禁止开发和生产自主武器系统的主要原因。

So far, the United States and other major militaries have made a commitment that a human will always be kept in the loop, and that specific authorization will be required before these machines engage in an attack that could result in loss of life. The reality, however, is that full automation on the battlefield will deliver enormous tactical advantages. No human can possibly react and make decisions at a speed comparable to that of artificial intelligence. Once one country breaches the current informal prohibition on full autonomy and begins to deploy these capabilities, it’s inevitable that any competing military would have to immediately follow suit or find itself at a critical disadvantage. This fear of falling behind is likely a major reason that the U.S., China and Russia are all opposed to a formal ban on the development and production of autonomous weapons systems.

我认为，通过观察另一种战争——华尔街人工智能交易系统之间的持续战斗，我们可以预见这一切将如何展开。算法交易现在主导着各大证券交易所的日常交易，占美国总交易量的 80%。早在 2013 年，一组物理学家就研究了金融市场，并在《自然》杂志上发表了一篇论文，宣称“一个以‘群体’掠夺性算法为特征的竞争机器的新兴生态”已经存在，算法交易可能已经发展到超出了设计这些系统的人类的控制范围，甚至超出了人类的理解范围。20^这些算法现在融合了人工智能的最新进展，它们对市场的影响力急剧增加，它们的互动方式变得更加难以理解。例如，许多算法能够直接利用彭博社和路透社等公司提供的机器可读新闻来源，然后在极短的时间内根据这些信息进行交易。对于短期瞬时交易，没有人能够理解正在发生的事情的细节，更不用说试图超越算法了。最终，我怀疑战场上发生的许多动态对抗也会如此。

I think we can get something of a preview of how all this might unfold by looking at another type of warfare—the continuous battle between AI-powered trading systems on Wall Street. Algorithmic trading now dominates the daily transactions on the major stock exchanges, accounting for as much as eighty percent of overall trading volume in the United States. As far back as 2013, a group of physicists studied financial markets and published a paper in the journal Nature declaring that “an emerging ecology of competitive machines featuring ‘crowds’ of predatory algorithms” existed and that algorithmic trading had perhaps already progressed beyond the control—and even comprehension—of the humans who designed the systems.²⁰ Those algorithms now incorporate the latest advances in AI, their influence on markets has increased dramatically, and the ways in which they interact have grown even more incomprehensible. Many algorithms, for example, have the ability to tap directly into machine-readable news sources provided by companies like Bloomberg and Reuters and then trade on that information in tiny fractions of a second. When it comes to short-term moment-by-moment trading, no human being can begin to comprehend the details of what is unfolding, let alone attempt to outsmart the algorithms. Eventually, I suspect the same will be true with many of the kinetic confrontations that occur on battlefields.

即使主要军队只部署自主战场技术，危险也是真实存在的。机器人战争的展开速度可能会超出军事或政治领导人完全了解或缓和局势的能力。换句话说，因相对较小的事件而陷入大战的风险可能会大大增加。另一个令人担忧的是，在一个机器人与机器人战斗、很少有人的生命受到直接威胁的世界里，人们认为发动战争的成本可能会变得低得令人不安。这在美国可能已经是一个问题，在美国，取消征兵制、采用全志愿军队，导致很少有社会精英送子女参军。因此，那些掌握最大权力的人几乎没有什么利益可言；他们往往不会直接承担军事行动的个人成本。我怀疑这种脱节可能对美国数十年来在中东的参与产生了重大影响。可以肯定的是，如果一台机器能够冒着危险保护士兵的生命，这无疑是一件好事。但我们需要非常小心，不要让这种低风险的认知影响我们在做出参战决定时的集体判断。

Even if autonomous battlefield technologies are deployed exclusively by major militaries, the dangers are very real. A robotic battle might unfold with a speed that could outstrip the ability of military or political leaders to fully understand or deescalate the situation. In other words, the risk of blundering into a major war as a result of a relatively minor incident might increase significantly. Another concern is that in a world where robots battle robots and very few human lives are immediately at stake, the perceived cost of going to war might become uncomfortably low. This is arguably already an issue in the United States, where the elimination of the draft in favor of an all-volunteer military has led to a situation in which very few of society’s elites send their children to serve in the armed forces. As a result, those who hold the most power have very little skin in the game; they tend to be insulated from the direct personal costs of military action. I suspect that this disconnect may have contributed significantly to the U.S.’s decades-long engagements in the Middle East. To be sure, if a machine can go into harm’s way and thereby preserve the life of a soldier, that is an unambiguously good thing. But we need to be quite careful that we don’t allow that perception of low risk to color our collective judgment when it comes to making a decision to go to war.

最大的危险是，一旦这些武器被生产出来，合法政府和军队可能无法保持对致命自主技术的控制。在这种情况下，这些武器最终可能会被非法军火商交易，将机枪或其他小型武器送到恐怖分子、雇佣兵或流氓国家手中。如果自主武器被广泛使用， 屠杀机器人视频中描述的噩梦场景很容易成为现实。即使这些武器无法购买，开发该技术的门槛也远低于其他大规模杀伤性武器。尤其是在无人机方面，同样易于获得的技术和部件，用于商业或娱乐应用，可能会被武器化。虽然制造核武器是一项巨大的挑战，即使一个民族国家的资源可用，但设计和部署一小群自主无人机可能只需要几个人在地下室工作就能完成。就像病毒一样，一旦自主武器技术进入环境，将很难防御或遏制，并可能引发混乱。

The greatest danger is that legitimate governments and militaries might fail to maintain control of lethal autonomous technology once the weapons were produced. In that event, the weapons could end up being traded by the types of illicit arms dealers that deliver machine guns or other small armaments into the hands of terrorists, mercenaries or rogue states. If autonomous weapons were to become widely available, the nightmare scenarios depicted in the Slaughterbots video could easily become reality. And even if the weapons were not available for purchase, the barriers to developing the technology are far lower than for other weapons of mass destruction. In the case of drones especially, the same easy-to-obtain technology and components intended for commercial or recreational applications could potentially be weaponized. While building a nuclear weapon poses an enormous challenge even when the resources of a nation-state are available, designing and deploying a small swarm of autonomous drones is something that might be accomplished by a few people working in a basement. Much like a virus, once autonomous weapons technology escapes into the environment, it will be very difficult to defend against or contain, and chaos may well ensue.

一个常见的错误（有时是媒体的唆使）是将致命自主武器的幽灵与我们在电影《终结者》等中看到的科幻场景混为一谈。这是一种分类错误，并且会分散我们对此类武器带来的近期危险的注意力。风险不在于机器会以某种方式摆脱我们的控制并决定自愿攻击我们。那将需要人工智能，正如我们所见，这至少在未来几十年内才会实现。相反，我们必须担心人类会选择如何处理这些并不比 iPhone 更“智能”的武器，但它们在识别、跟踪和杀死目标方面却无情地表现出色。这绝不是一个纯粹的未来主义问题。正如 Stuart Russell 在Slaughterbots视频的结尾所说，这部电影戏剧化地展现了“集成和小型化我们已有技术的结果”。换句话说，这些武器很有可能在未来几年内出现，如果我们想阻止这种情况发生，“采取行动的时间窗口正在迅速关闭。” ²¹鉴于联合国批准的全面禁止此类武器的禁令可能不会很快到来，国际社会至少应该集中精力确保此类武器永远不会落入恐怖分子或其他非国家行为者手中，并可能将其用于对付平民。

One common mistake—sometimes abetted by the media—is to conflate the specter of lethal autonomous weapons with the science fiction scenarios we’ve all seen in movies like The Terminator. This is a category error—and a dangerous distraction from the near-term dangers posed by such weapons. The risk is not that the machines will somehow wrest themselves from our control and decide to attack us of their own volition. That would require artificial general intelligence, which as we have seen, likely lies at least decades in the future. Rather, we have to worry about what human beings will choose to do with weapons that are no more “intelligent” than an iPhone—but which are ruthlessly competent at identifying, tracking and killing targets. And this is by no means a purely futuristic concern. As Stuart Russell says in the conclusion of the Slaughterbots video, the film dramatizes “the results of integrating and miniaturizing technologies that we already have.” In other words, these are weapons that could quite possibly emerge with the next few years, and if we want to prevent that, “the window to act is closing fast.”²¹ Given that an outright United Nations–sanctioned ban on the weapon may not be coming anytime soon, the international community should at a minimum focus on ensuring that such weapons never become accessible to terrorists or other non-state actors who might deploy them against civilians.

机器学习算法中的偏见、公平性和透明度

BIAS, FAIRNESS AND TRANSPARENCY IN MACHINE LEARNING ALGORITHMS

随着人工智能和机器学习的应用越来越广泛，至关重要的是，这些算法产生的结果和建议必须公平，并且其背后的原因必须得到充分解释。如果你使用深度学习系统来最大限度地提高某些工业机器的能源效率，那么你可能并不特别关心推动算法结果的细节；你只是想要最佳结果。但是，当机器学习应用于刑事司法、招聘决策或住房抵押贷款申请处理等领域时——换句话说，应用于直接影响人类权利和未来福祉的高风险决策时，至关重要的是，算法结果必须能够证明对各个人口群体不带偏见，并且导致这些结果的分析必须透明和公正。

As artificial intelligence and machine learning are deployed more and more widely, it’s critical that the results and recommendations produced by these algorithms are perceived as fair and that the reasoning behind them can be adequately explained. If you’re using a deep learning system to maximize the energy efficiency of some industrial machine, then you are probably not particularly concerned about the details that drive an algorithmic outcome; you simply want the optimal result. But when machine learning is applied to areas like criminal justice, hiring decisions or the processing of home mortgage applications—in other words, to high-stakes decisions that directly impact the rights and future well-being of human beings—it’s essential that algorithmic outcomes can be shown to be unbiased across demographic groups and that the analysis that led to those outcomes is transparent and just.

偏见是机器学习中常见的问题，在大多数情况下，问题的出现是因为用于训练算法的数据存在问题。正如我们在上一章中看到的，西方开发的面部识别算法往往对有色人种存在偏见，因为训练数据集往往包含绝大多数白人面孔。一个更普遍的问题是，用于训练算法的大部分数据直接来自人类的行为、决策和行动。如果生成数据的人在某种程度上存在偏见（例如基于种族或性别），那么这种偏见将自动封装在训练数据集中。

Bias is a common issue in machine learning, and in most cases the problem arises because of problems with the data used to train the algorithms. As we saw in the previous chapter, facial recognition algorithms developed in the West are often biased against people of color because the training dataset tends to contain overwhelmingly white faces. A more general problem is that much of the data that is used to train algorithms results directly from human behaviors, decisions and actions. If the humans who generate the data are biased in some way—for example on the basis of race or gender—then that bias will automatically be encapsulated in the training dataset.

举个例子，考虑一个机器学习算法，该算法旨在筛选大型公司空缺职位的简历。这样的系统可能会根据过去申请类似职位的求职者提交的所有简历全文以及招聘经理对每份简历做出的决定进行训练。机器学习算法会仔细研究所有这些数据，并了解简历的特征，这些特征可能会导致决定让求职者参加进一步的面试，以及表明应拒绝该求职者而不作进一步考虑的简历属性。当人力资源部门必须筛选数百或数千名潜在求职者时，能够有效地做到这一点并生成可管理的顶级候选人名单的算法可能会节省大量时间，因此，像这样的简历筛选系统正变得越来越流行，尤其是在大型公司中。然而，假设算法所依据的过去的招聘决定反映了招聘经理某种程度的公开或潜意识的种族主义或性别歧视。在这种情况下，机器学习系统会在正常训练过程中自动拾取这种偏见。算法的创建者并无恶意；偏见存在于训练数据中。结果将是一个延续甚至放大现有人类偏见的系统，并且对有色人种或女性明显不公平。

As an example, consider a machine learning algorithm designed to screen resumes for an open job at a large corporation. Such a system might be trained on the full text of all the resumes received from past applicants to similar jobs, together with the decisions that hiring managers made for each of those resumes. The machine learning algorithm would churn through all this data and assimilate an understanding of the characteristics of a resume that will likely lead to a decision to bring a job candidate in for further interviews, as well as resume attributes that suggest the applicant should be rejected without further consideration. An algorithm that can do this effectively, generating a manageable list of top-ranked candidates, is likely to save a great deal of time when a human resource department has to weed through hundreds or thousands of potential applicants, and for this reason, resume screening systems like this are becoming popular, especially in large companies. Suppose, however, that the past hiring decisions upon which the algorithm is trained reflect some degree of overt or subconscious racism or sexism on the part of hiring managers. In that case, the machine learning system will automatically pick up that bias as it goes through its normal training process. There is no nefarious intent on the part of the creators of the algorithm; the bias exists in the training data. The result would be a system that perpetuated, or perhaps even amplified, existing human biases and would be demonstrably unfair to people of color or women.

2018 年，亚马逊也发生过类似的事情，当时该公司停止了机器学习系统的开发，因为事实证明该系统在筛选技术职位简历时对女性存在偏见。事实证明，当简历中包含“女性”一词时（例如在提到女性俱乐部或运动时，或者当应聘者毕业于女子大学时），系统会给简历打较低的分数，使女性求职者处于不利地位。即使亚马逊的开发人员针对发现的具体问题进行了更正，也无法保证算法不会出现偏见，因为其他变量可能充当了性别的代理。22^值得注意的是，这并不一定意味着之前的招聘决策存在彻底的性别歧视。算法可能被训练成带有偏见，仅仅是因为女性在技术岗位上的代表性不足，因此男性占了绝大多数。据亚马逊称，该算法从未超出开发阶段，也从未真正用于筛选简历，但如果该算法被部署，毫无疑问会巩固女性在技术岗位中代表性不足的现状。

Something very similar to this happened at Amazon in 2018 when the company halted development of a machine learning system because it was shown to be biased against women when screening resumes for technical positions. It turned out that when a resume included the word “women’s,” as might happen in a reference to women’s clubs or sports or when the candidate has graduated from an all-women’s college, the system gave the resume a lower score, putting the female job candidate at a disadvantage. Even when Amazon’s developers made corrections for the specific problems it discovered, it wasn’t possible to guarantee that the algorithm would be unbiased because other variables might be acting as a proxy for gender.²² It’s important to note that this did not necessarily imply outright sexism in prior hiring decisions. The algorithm may have been trained to be biased simply because women are underrepresented in technical roles, and therefore men constitute the vast majority of hires. According to Amazon, the algorithm never made it beyond the development phase and was never actually used to screen resumes, but if it had been deployed, it would unquestionably have worked to solidify the underrepresentation of women in technical jobs.

当机器学习系统用于刑事司法系统时，风险会更高。此类算法通常用于协助做出保释、假释或量刑决定。其中一些系统由州或地方政府开发，而另一些则由私人公司设计和销售。2016 年 5 月， Propublica发表了一篇关于 COMPAS 算法的分析，该算法被广泛用于预测某个人在获释后成为重复犯罪者的可能性。23^分析表明，非裔美国被告被不公平地分配了比白人被告更高的风险。这一评估似乎得到了传闻证据的支持。Propublica的文章中有一个故事，一名 18 岁的黑人女性骑着一辆对她来说太小的儿童自行车走了一小段路，然后在车主反对后放弃了它。换句话说，这更像是一种恶作剧行为，而不是严重的盗窃企图。尽管如此，这名年轻女子还是被捕了，当她被关进监狱等待出庭时，COMPAS 系统被应用于她的案件。结果显示^，算法判断出她成为累犯的风险明显高于一名 41 岁的白人男子，该男子此前曾因持械入室盗窃被判刑 5 年。24销售 COMPAS 系统的 Northpoint, Inc. 对Propublica的分析提出异议，而关于该系统到底有多大程度存在偏见的争论也一直存在。然而，尤其令人担忧的是，该公司不愿分享其算法的计算细节，因为它认为这些信息是专有的。换句话说，第三方无法对系统进行详细审核，以检查是否存在偏见或准确性。显然，在利用算法做出对人类生命至关重要的决策时，需要有更多的透明度和监督。

An even higher stakes situation occurs when machine learning systems are used in the criminal justice system. Such algorithms are often used to assist with making decisions on bail, parole or sentencing. Some of these systems are developed by state or local governments, while others are designed and sold by private companies. In May 2016, Propublica published an analysis of an algorithm called COMPAS that is widely used to predict the likelihood that a particular individual will be a repeat offender upon release.²³ The analysis suggested that African American defendants were unfairly being assigned a higher risk than white defendants. That assessment seemed to be supported by anecdotal evidence. Propublica’s article included the story of an eighteen-year-old black woman who rode a child’s bicycle that was far too small for her for a short distance before abandoning it after the owner objected. In other words, something that seems more like an instance of mischievous behavior than a serious attempt at theft. The young woman was nonetheless arrested, and the COMPAS system was applied to her case when she was booked into jail to await a court appearance. It turned out that the algorithm assigned her a significantly higher risk of becoming a repeat offender than a forty-one-year-old white man who already had a prior conviction for armed burglary and had served five years in prison.²⁴ The company that sells the COMPAS system, Northpoint, Inc., disputes the analysis performed by Propublica, and there continues to be a debate about the extent to which the system is actually biased. It is especially concerning, however, that the company is unwilling to share the computational details of its algorithm because it considers them to be proprietary. In other words, there is no way for a third party to perform a detailed audit of the system for bias or accuracy. It seems clear that when algorithms are deployed to make decisions that are so extraordinarily consequential for human lives, there needs to be more transparency and oversight.

虽然训练数据中的偏见是机器学习系统不公平现象最常见的原因，但它并不是唯一起作用的因素。算法本身的设计也会引入或放大偏见。例如，假设一个面部识别系统是在与美国人口分布完全一致的数据集上进行训练的。由于非裔美国人只占总人口的 13%，因此该系统最终仍可能对黑人产生偏见。这个问题的严重程度（是加剧了还是减轻了）将取决于算法设计中的技术决策。

Though bias in the training data is the most common cause of unfairness in machine learning systems, it is not the only factor at play. The design of the algorithms themselves can also introduce or amplify bias. For example, suppose a facial recognition system was trained on a dataset that exactly mirrored the demographic distribution of the U.S. population. Because African Americans are only about thirteen percent of the population, the system could still end up biased against black people. The extent to which this became an issue—whether the problem was amplified or mitigated—would be determined by technical decisions made in the design of the algorithm.

好消息是，设计公平透明的机器学习系统已成为人工智能研究的主要重点。所有大型科技公司都在这一领域投入了大量资金。谷歌、Facebook、微软和 IBM 都发布了旨在帮助开发人员在机器学习算法中建立公平性的软件工具。使深度学习系统可解释且透明，以便可以审核结果是一个特殊的问题，因为深度神经网络往往是一种“黑匣子”，其中对输入数据的分析和理解分布在数百万个人工神经元之间的连接中。同样，评估和确保公平是一个非常具有挑战性且技术性很强的问题。正如亚马逊在其简历筛选系统中发现的那样，简单地调整算法以忽略种族或性别等参数并不是一个合适的解决方案，因为系统可能会关注代理。例如，求职者的名字可能表示性别，而街区或邮政编码可能是种族的代理。一种特别有希望实现人工智能公平性的方法是使用反事实。借助这项技术，可以检查系统是否在种族、性别或性取向等敏感变量更改为不同值时产生相同的结果。不过，这些领域的研究才刚刚起步，还需要做更多的工作才能开发出能够始终如一地产生真正公平的机器学习系统的技术。

The good news is that designing machine learning systems to be fair and transparent has become a major focus of AI research. All the major tech companies are making significant investments in this area. Google, Facebook, Microsoft and IBM have all released software tools designed to help developers build fairness into machine learning algorithms. Making deep learning systems explainable and transparent so that outcomes can be audited is a particular problem because deep neural networks tend to be a kind of “black box,” in which analysis and comprehension of the input data is distributed across millions of connections between artificial neurons. Likewise, assessing and ensuring fairness is a very challenging and highly technical issue. As Amazon found with its resume screening system, simply tweaking the algorithm to ignore parameters like race or gender is not an adequate solution because the system might instead focus on proxies. For example, the first name of a job candidate might indicate gender, and the neighborhood or zip code could be a proxy for race. One especially promising approach to AI fairness is the use of counterfactuals. With this technique, a system is checked to verify that it produces the same outcome when sensitive variables like race, gender or sexual orientation are changed to different values. Still, research in these areas is just getting started, and it will take a lot more work to develop techniques that will consistently result in machine learning systems that are truly fair.

人工智能在高风险决策中的最终前景是，它能够可靠地减少偏见，提高准确性，而不像人类那样仅凭判断。虽然修复算法中的偏见可能具有挑战性，但它几乎总是比人类更容易。正如麦肯锡全球研究院主席詹姆斯·曼尼卡告诉我的那样，“一方面，机器系统可以帮助我们克服人类的偏见和易犯错误，但另一方面，它们也可能带来更大的问题。” ²⁵尽量减少或消除这些公平问题是人工智能领域面临的最关键和最紧迫的挑战之一。

The ultimate promise of AI deployed in high-stakes decisions is a technology that reliably produces less bias and greater accuracy than human judgment alone. Though fixing bias in an algorithm can be challenging, it is nearly always much easier than doing the same for a human being. As McKinsey Global Institute Chairman James Manyika told me, “On the one hand, machine systems can help us overcome human bias and fallibility, and yet on the other hand, they could also introduce potentially larger issues of their own.”²⁵ Minimizing or eliminating those fairness issues is one of the most critical and urgent challenges facing the field of artificial intelligence.

为了实现这一结果，构建、测试和部署人工智能算法的开发人员来自不同的背景也很重要。鉴于人工智能将塑造我们的经济和社会，最了解这项技术并因此最有能力影响其方向的专家必须代表整个社会。然而，实现这一目标的进展迄今为止有限。2018 年的一项研究发现，女性仅占领先人工智能研究人员的 12% 左右，而代表性不足的少数群体的比例甚至更低。正如斯坦福大学的李飞飞所说：“如果我们环顾四周，无论是公司里的人工智能团队、学术界的人工智能教授、人工智能博士生还是顶级人工智能会议上的人工智能演讲者，无论你从哪个角度看：我们缺乏多样性。我们缺乏女性，我们缺乏代表性不足的少数群体。” ²⁶ 大学、大型科技公司和几乎所有顶尖的人工智能研究人员都坚定地致力于改变这一现状。李飞飞与他人共同创办了一项特别有前途的倡议：AI4ALL 是一个致力于通过为有才华的高中生提供夏令营来吸引年轻女性和弱势群体进入人工智能领域的组织。该组织发展迅速，目前在美国 11 所大学提供暑期项目。虽然还有很多工作要做，但像 AI4ALL 这样的项目加上行业对吸引包容性人工智能人才的承诺，很可能在未来几年和几十年内产生出更加多元化的研究人员群体。为该领域带来更广泛的视角，很可能直接转化为更有效、更公平的人工智能系统。

In order to achieve this outcome, it’s also important that the developers building, testing and deploying AI algorithms come from diverse backgrounds. Given that artificial intelligence is poised to shape our economy and society, it is essential that the experts who best understand the technology—and are therefore best positioned to influence its direction—are representative of society as a whole. Progress in reaching this goal, however, has so far been limited. A 2018 study found that women represent only about twelve percent of leading artificial intelligence researchers, and the numbers for underrepresented minorities are even lower. As Stanford’s Fei-Fei Li says, “If we look around, whether you’re looking at AI groups in companies, AI professors in academia, AI PhD students or AI presenters at top AI conferences, no matter where you cut it: we lack diversity. We lack women, and we lack under-represented minorities.”²⁶ The universities, the major tech companies and nearly all top AI researchers are firmly committed to changing this. One especially promising initiative was co-founded by Li: AI4ALL is an organization dedicated to attracting young women and underrepresented groups into the field of artificial intelligence by providing summer camps for talented high school students. The organization has expanded rapidly and now offers summer programs on eleven university campuses in the United States. While much work remains to be done, programs like AI4ALL together with an industry commitment to attracting inclusive AI talent will likely produce a significantly more diverse set of researchers in the coming years and decades. Bringing a broader range of perspective into the field will likely translate directly into more effective and fair artificial intelligence systems.

超级智能带来的生存威胁和“控制问题”

AN EXISTENTIAL THREAT FROM SUPERINTELLIGENCE AND THE “CONTROL PROBLEM”

人工智能的最大风险在于，拥有超人智能的机器有朝一日可能会摆脱人类的直接控制，并采取最终对人类生存构成威胁的行动。安全问题、武器化和算法偏见都构成了直接或短期的危险。这些担忧显然需要我们现在就解决——以免为时已晚。然而，超级智能带来的生存威胁则更具推测性，而且几乎肯定要等到几十年后——甚至一个世纪或更久以后。尽管如此，这种风险还是吸引了许多知名人士的关注，并得到了媒体的大量炒作和关注。

The AI risk that transcends all others is the possibility that machines with superhuman intelligence might someday wrest themselves from our direct control and pursue a course of action that ultimately presents an existential threat to humanity. Security issues, weaponization and algorithmic bias all pose immediate or near-term dangers. These are concerns that we clearly need to be addressing right now—before it is too late. An existential threat from superintelligence, however, is far more speculative and almost certainly lies decades—or perhaps even a century or more—in the future. Nonetheless, it is this risk that has captured the imagination of many prominent people and has received an enormous amount of media hype and attention.

2014 年，人工智能存在的风险成为公众严肃讨论的话题。同年 5 月，包括剑桥大学宇宙学家斯蒂芬·霍金、人工智能专家 Stuart Russell、物理学家 Max Tegmark 和 Frank Wilczek 在内的一组科学家共同撰写了一封公开信，发表在英国 《独立报》上，宣称人工智能超级智能的出现“将成为人类历史上最大的事件”，具有超人智力的计算机可能能够“智胜金融市场、超越人类研究人员的发明能力、超越人类领导者的操纵能力，并开发出我们甚至无法理解的武器”。这封信警告说，不认真对待这一迫在眉睫的危险很可能成为人类“历史上最严重的错误” ^。27

The specter of existential AI risk emerged as a topic of serious public discussion in 2014. In May of that year, a group of scientists including the University of Cambridge cosmologist Stephen Hawking along with AI expert Stuart Russell and physicists Max Tegmark and Frank Wilczek co-authored an open letter published in the U.K.’s Independent declaring that the advent of artificial superintelligence “would be the biggest event in human history,” and that a computer with superhuman intellectual capability might be capable of “outsmarting financial markets, out-inventing human researchers, out-manipulating human leaders, and developing weapons we cannot even understand.” The letter warned that a failure to take this looming danger seriously might well turn out to be humanity’s “worst mistake in history.”²⁷

同年晚些时候，牛津大学哲学家尼克·博斯特罗姆出版了他的书《超级智能：路径、危险、策略》，这本书很快就成为了一本令人惊讶的畅销书。博斯特罗姆在书的开篇指出，人类统治地球纯粹是基于优越的智力。许多其他动物比我们更快、更强壮、更凶猛；是我们的大脑让我们占据了主导地位。一旦另一个实体的智力大大超过我们，局面就很容易被扭转。正如博斯特罗姆所说，“就像大猩猩的命运现在更多地取决于我们人类而不是大猩猩本身一样，我们物种的命运也将取决于机器超级智能的行为。” ²⁸

Later that same year, the Oxford University philosopher Nick Bostrom published his book Superintelligence: Paths, Dangers, Strategies, which quickly became a somewhat surprising bestseller. Bostrom opens the book by pointing out that humans rule the earth purely on the basis of superior intellect. Many other animals are faster, stronger or more ferocious; it is our brains that led to dominance. Once another entity dramatically exceeds our own intellectual capability, the tables could easily be turned. As Bostrom puts it, “just as the fate of gorillas now depends more on us humans than on the gorillas themselves, so the fate of our species would depend on the actions of the machine superintelligence.”²⁸

博斯特罗姆的书影响巨大，尤其是在硅谷精英阶层。在这本书出版后一个月内，埃隆·马斯克就宣称“有了人工智能，我们就是在召唤恶魔”，而且人工智能“可能比核武器更危险”。29 一年后^，马斯克与他人共同创立了 OpenAI，并赋予其构建“友好”人工智能的具体使命。在那些最受博斯特罗姆观点影响的人中，人工智能有朝一日会构成生存威胁的想法开始被视为几乎肯定的事情——而且这种危险最终远比气候变化或全球流行病等更平凡的问题更为可怕和严重。在一次观看次数超过 500 万次的 Ted Talk 中，神经科学家和哲学家山姆·哈里斯认为“很难想象 [我们在人工智能方面取得的进步] 不会摧毁我们或激励我们自我毁灭”，并建议“我们需要类似曼哈顿计划的东西”，专注于通过弄清楚如何构建友好、可控的人工智能来避免这种结果^。30

Bostrom’s book was enormously influential, especially among the Silicon Valley elite. Within a month of its publication, Elon Musk was declaring that “with artificial intelligence, we are summoning the demon” and that AI “could be more dangerous than nuclear weapons.”²⁹ A year later, Musk would co-found OpenAI and give it the specific mission of building “friendly” artificial intelligence. Among those most deeply influenced by Bostrom’s arguments, the idea that AI will someday pose an existential threat began to be perceived as a near certainly—and a danger ultimately far more terrifying and consequential than more mundane concerns like climate change or global pandemics. In a Ted Talk with more than five million views, the neuroscientist and philosopher Sam Harris argues that “it’s very difficult to see how [the gains we make in artificial intelligence] won’t destroy us or inspire us to destroy ourselves” and suggests that “we need something like a Manhattan Project” focused on avoiding that outcome by figuring out how to build friendly, controllable AI.³⁰

当然，在我们成功建造出一台认知能力至少与我们相当的真正能思考的机器之前，所有这些都不是问题。正如我们在第 5 章中看到的那样，通向通用人工智能的道路上还有无数的重大障碍，可能需要几十年的时间才能实现达到这一里程碑所必需的突破。回想一下，我为撰写《智能建筑师》一书而采访过的顶尖人工智能研究人员平均估计通用人工智能的到来需要大约80 年——或者本世纪末。然而，一旦人类水平的人工智能成为现实，超级智能几乎肯定会迅速跟进。事实上，任何具有与人类水平一样学习和推理能力的机器智能都将比我们更优秀，因为它还将享受计算机已经拥有的所有优势——包括以难以理解的速度计算和操控信息的能力，以及直接通过网络与其他机器交互的能力。

None of this will be a concern, of course, until we manage to build a true thinking machine with cognitive capability at least equivalent to our own. As we saw in Chapter 5, the path to artificial general intelligence contains an unknown number of major hurdles, and it will likely take decades to achieve the necessary breakthroughs to reach this milestone. Recall that the mean estimate for AGI arrival from the top AI researchers I spoke to for my book Architects of Intelligence was about eighty years—or the end of this century. Once human-level AI becomes a reality, however, it is almost certain that superintelligence will rapidly follow. Indeed, any machine intelligence with the ability to learn and reason at the level of a human being would already be superior to us simply because it would also enjoy all the advantages that computers already have over us—including the ability to calculate and manipulate information at incomprehensible speed and to directly interface with other machines across networks.

除此之外，大多数人工智能专家认为，这种机器智能很快就会决定将其智力能量用于改进自己的设计。随着系统变得越来越聪明，越来越善于重新设计自己的人工智能，这将导致持续不断的递归改进。其结果必然是“智能爆炸”——雷·库兹韦尔等技术乐观主义者认为，这一现象将成为奇点和新时代黎明的催化剂。早在摩尔定律开始提供可能将这种事件带入可能性领域的计算机硬件之前，人工智能的进步将有朝一日导致机器智能爆炸的论点就已经形成。1964 年，数学家 IJ Good 撰写了一篇题为“关于第一台超智能机器的推测”的学术论文，他在论文中这样解释这一概念：

Beyond this point, most AI experts assume that such a machine intelligence would soon decide to turn its intellectual energy toward improving its own design. This would then lead to relentless, recursive improvement as the system became ever smarter and more adept at re-engineering its own artificial mind. The result would inevitably be an “intelligence explosion”—a phenomenon that technoptimists like Ray Kurzweil believe will be the catalyst for the Singularity and the dawn of a new age. The argument that advances in AI would someday produce an explosion in machine intelligence was formulated long before Moore’s Law began to deliver computer hardware that might bring such an event into the realm of possibility. In 1964, the mathematician I. J. Good wrote an academic paper entitled “Speculations Concerning the First Ultraintelligent Machine” in which he explained the concept like this:

让超智能机器被定义为能够远远超越任何聪明人的所有智力活动的机器。由于机器设计是这些智力活动之一，超智能机器可以设计出更好的机器；毫无疑问，这将带来“智能爆炸”，而人类的智能将远远落后。因此，第一台超智能机器是人类需要制造的最后一项发明，前提是这台机器足够温顺，可以告诉我们如何控制它^。31

Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an “intelligence explosion,” and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.³¹

超级智能机器将成为我们最后需要的发明，这一承诺体现了奇点支持者的乐观态度。机器必须保持足够温顺才能被控制，这一条件表明存在存在威胁的可能性。超级智能的这一阴暗面在人工智能界被称为“控制问题”或“价值观一致问题”。

The promise that a superintelligent machine would be the last invention we ever need to make captures the optimism of Singularity proponents. The qualification that the machine must remain docile enough to be kept under control is the concern that suggests the possibility of an existential threat. This dark side of superintelligence is known in the AI community as the “control problem” or the “value alignment problem.”

控制问题并非源于对《终结者》等电影中描述的那种公然恶意的机器的恐惧。每个人工智能系统都是围绕一个目标函数设计的，换句话说，就是系统努力实现的一个特定目标（用数学术语表示）。令人担忧的是，超级智能系统在设定了这样的目标后，可能会不遗余力地使用一些手段来追求这一目标，而这些手段可能会产生意想不到的后果，对我们的文明造成损害甚至致命。一个涉及“回形针最大化器”的思想实验经常被用来说明这一点。想象一个超级智能，其特定目标是优化回形针生产。在不遗余力地追求这一目标的过程中，超级智能机器可能会发明新技术，使它能够将地球上几乎所有的资源转化为回形针。由于该系统的智力远远超出我们，因此它很可能能够成功挫败任何试图关闭它或改变其行动方针的企图。事实上，任何干扰企图都会与系统的目标函数相悖，而且系统有明显的动机去阻止这种行为。

The control problem is not driven by fear of overtly malevolent machines of the kind portrayed in movies like The Terminator. Every AI system is designed around an objective function, in other words, a specific goal, expressed in mathematical terms, that the system strives to achieve. The concern is that a superintelligent system, given such an objective, might relentlessly pursue it using means that have unintended or unanticipated consequences that could turn out to be detrimental or even fatal to our civilization. A thought experiment involving a “paperclip maximizer” is often used to illustrate this point. Imagine a superintelligence designed with the specific objective of optimizing paperclip production. As it relentlessly pursued this goal, a superintelligent machine might invent new technologies that would allow it to convert virtually all the resources on earth into paperclips. Because the system would be so far beyond us in terms of its intellectual capability, it would likely be able to successfully foil any attempt to shut it down or alter its course of action. Indeed, any attempt at interference would be at odds with the system’s objective function, and it would have a clear incentive to prevent this.

这个例子显然是想画个漫画。未来可能出现的真实情况可能更加微妙，而潜在的后果将更加难以预测，甚至不可能预测。我们已经可以举出一个重要的例子，说明意外后果显然会损害社会结构。YouTube 和 Facebook 等科技公司使用的机器学习算法通常被赋予最大化平台用户参与度的目标。这反过来又会带来更多的在线广告收入。然而，很明显，追求这一目标的算法很快就发现，让人们参与进来的最佳方式是向他们提供更多政治两极化的内容，或者直接利用愤怒或恐惧等情绪。例如，这导致了 YouTube 上经常出现的“兔子洞”现象，在适度的视频之后，会连续推荐更极端的内容，所有这些都会导致人们持续地以情感为驱动与平台互动。32^这可能有利于盈利，但显然不利于我们的社会或政治环境。如果超级智能系统也出现类似的误判，那么它在追求目标的过程中很可能无法重新获得控制权。

This example is obviously intended as a kind of cartoon. The real scenarios that might unfold in the future would likely be far more subtle, and the potential consequences would be much more difficult—or perhaps impossible—to anticipate in advance. We can already point to one important example that illustrates how unintended consequences can clearly be detrimental to the social fabric. The machine learning algorithms utilized by tech companies like YouTube and Facebook have generally been given the objective of maximizing user engagement on the platform. This in turn leads to more revenue from online advertisements. However, it has become evident that the algorithms pursuing this objective soon figured out that the best way to keep people engaged is to feed them ever more politically polarized content or tap directly into emotions like outrage or fear. This, for example, has led to the often-noted “rabbit hole” phenomenon on YouTube, in which a moderate video is followed by successive recommendations for ever more extreme content, all of which leads to sustained emotion-driven engagement with the platform.³² That may be good for profitability, but it’s clearly not good for our social or political environment. If a similar miscalculation were made with a superintelligent system, it might well be impossible to regain control as it sought to pursue its objective.

寻找控制问题的解决方案已成为大学学术研究的重要课题，尤其是在专门的私人资助组织中，例如 OpenAI、由尼克·博斯特罗姆 (Nick Bostrom) 领导的牛津大学人类未来研究所和位于加利福尼亚州伯克利的机器智能研究所。斯图尔特·罗素 (Stuart Russell) 在其 2019 年出版的《与 人类兼容：人工智能与控制问题》一书中指出，解决这个问题的最佳方法是根本不要在高级人工智能系统中构建明确的目标函数。相反，系统应该被设计为“最大限度地实现人类的偏好”。33^因为机器智能永远无法确定这些偏好或意图是什么，所以它必须通过研究人类行为来制定目标，并愿意与人类对话并接受人类的指导。与势不可挡的回形针最大化器不同，如果这样的系统认为这符合它旨在优化的人类偏好，它就会屈服于被关闭。

The quest to find a solution to the control problem has become an important topic of academic research at universities and especially within specialized, privately funded organizations such as OpenAI, Oxford University’s Future of Humanity Institute, which is directed by Nick Bostrom, and the Machine Intelligence Research Institute located in Berkeley, California. In his 2019 book Human Compatible: Artificial Intelligence and the Problem of Control, Stuart Russell argues that the best solution to the problem is to not build an explicit objective function into advanced AI systems at all. Instead, systems should be designed to “maximize the realization of human preferences.”³³ Because the machine intelligence could never be certain what these preferences or intentions are, it would have to formulate its objectives by studying human behavior and would be willing to dialogue with and accept guidance from humans. Unlike the unstoppable paperclip maximizer, such a system would submit to being shut down if it believed this was in line with the human preferences it was designed to optimize.

这与当前构建 AI 系统的方法截然不同。正如 Russell 所解释的那样：

This represents a stark departure from the current approach to building AI systems. As Russell explains:

实际上，将这样的模型付诸实践需要进行大量研究。我们需要“微创”的决策算法，防止机器干扰它们不确定其价值的部分世界，以及让机器更多地了解我们对未来应该如何展开的真实、潜在偏好。然后，这样的机器将面临一个古老的道德哲学问题：如何在具有冲突愿望的不同个体之间分配利益和成本。

Actually putting a model like this into practice requires a great deal of research. We need “minimally invasive” algorithms for decision making that prevent machines from messing with parts of the world whose value they are unsure about, as well as machines that learn more about our true, underlying preferences for how the future should unfold. Such machines will then face an age-old problem of moral philosophy: how to apportion benefits and costs among different individuals with conflicting desires.

这一切可能需要十年才能完成——即便如此，仍需要制定法规来确保采用可证明安全的系统，而不符合要求的系统则被淘汰。这并不容易。但很明显，在人工智能系统的能力在重要领域超越人类之前，必须建立这种模式^。34

All this could take a decade to complete—and even then, regulations will be required to ensure provably safe systems are adopted while those that don’t conform are retired. This won’t be easy. But it’s clear that this model must be in place before the abilities of A.I. systems exceed those of humans in the areas that matter.³⁴

值得注意的是，除了大学顶尖人工智能教科书的合著者斯图尔特·拉塞尔 (Stuart Russell) 之外，几乎所有警告超级智能可能存在生存威胁的知名声音都来自人工智能研究或计算机科学领域之外。敲响警钟的主要是萨姆·哈里斯这样的公共知识分子、马斯克这样的硅谷巨头，或者霍金或麻省理工学院物理学家马克斯·泰格马克等其他领域的科学家。大多数从事人工智能实际研究的专家往往更加乐观。当我为我的书《智能建筑师》采访了 23 位精英研究人员时，我发现虽然少数人认真对待生存威胁的可能性，但绝大多数人对此不以为然。一种普遍的看法是，超级智能的出现还很遥远，而要解决问题的具体参数又很模糊，因此没有必要继续研究这个问题。曾领导谷歌和百度人工智能研究团队的吴恩达（Andrew Ng）曾说过一句著名的话：担心人工智能的生存威胁就像担心火星人口过剩一样——甚至早在第一批宇航员被派往火星之前。机器人专家罗德尼·布鲁克斯（Rodney Brooks）也表达了同样的看法，他说超级智能还遥不可及，“它不会是和现在一模一样的世界，而是中间夹着一个人工智能超级智能……我们根本不知道这个世界或[超级智能 AI 系统]会是什么样子。预测人工智能的未来只是那些生活在远离现实世界的孤立学者的权力游戏。这并不是说这些技术不会到来，而是在它们到来之前我们无法知道它们会是什么样子。” ³⁵

It’s notable that with the exception of Stuart Russell, who is a co-author of the leading university artificial intelligence textbook, nearly all of the most prominent voices warning of a potential existential threat come from outside the fields of AI research or computer science. Instead, the alarm is primarily being sounded by public intellectuals like Sam Harris, Silicon Valley titans like Musk or scientists in other fields like Hawking or the MIT physicist Max Tegmark. Most of the experts engaged in actual AI research tend to be more sanguine. When I interviewed twenty-three elite researchers for my book Architects of Intelligence, I found that while a few took the possibility of an existential threat seriously, the vast majority were quite dismissive. A common refrain is that the emergence of superintelligence is so far off, and the specific parameters of the problem to be solved so nebulous, that there’s little point in pursuing the issue. Andrew Ng, who led AI research groups at Google and Baidu, is famous for saying that worrying about an existential threat from AI is like worrying about overpopulation on Mars—long before even the first team of astronauts has been sent to the red planet. The roboticist Rodney Brooks echoes this sentiment, saying that superintelligence is so far in the future that “it’s not going to be a case of having exactly the same world as it is today, but with an AI superintelligence in the middle of it.… We have no clue at all about what the world or [a superintelligent AI system] are going to be like. Predicting an AI future is just a power game for isolated academics who live in a bubble away from the real world. That’s not to say that these technologies aren’t coming, but we won’t know what they will look like before they arrive.”³⁵

提倡严肃对待人工智能威胁的人士强烈反对这样一种观点，即这个问题不重要或无法解决，因为它可能要几十年后才会出现。他们指出，控制问题需要在第一个超级智能出现之前解决，否则就太晚了。斯图尔特·拉塞尔喜欢用外星人的到来来打比方。想象一下，我们收到来自太空的信号，宣布外星人将在五十年后到达这里。想必我们会立即采取重大的全球行动来为这一事件做准备。拉塞尔认为，我们应该为超级智能的最终到来做同样的事情。

Advocates for taking an existential AI threat seriously push back strongly against the idea that the issue is unimportant or unapproachable simply because it likely will not arise until decades have passed. They point out that the control problem needs to be solved before the first superintelligence comes into existence—or it will be too late. Stuart Russell likes to make an analogy to the arrival of extraterrestrials. Imagine we received a signal from space announcing that the aliens will be here in fifty years. Presumably we would immediately put in place a major global effort to prepare for the event. Russell believes we should be doing the same for the eventual arrival of superintelligence.

我个人认为，应该认真对待人工智能可能带来的威胁。我认为，未来人类研究所等组织的研究人员正在积极研究这个问题，这是一件非常积极的事情。然而，在我看来，这代表了资源的合理分配，而且至少目前，这个问题最好在安静的学术研究环境中解决。目前，很难证明任何像政府资助的“曼哈顿计划”规模的计划是合理的。试图将这个问题注入已经失灵的政治进程似乎也不明智。我们真的希望对这项技术知之甚少或一无所知的政客在推特上谈论超级智能机器的危险吗？鉴于美国政府的能力非常有限，几乎无法完成任何事情，我还担心，炒作或政治化未来的生存威胁会分散人们对非常现实和直接的人工智能风险的注意力——包括武器化、安全和偏见——我们现在确实需要开始投入大量资源来解决这些问题。

My own view is that the potential for an existential AI threat should be taken seriously. I think it’s a very positive thing that researchers at organizations like the Future of Humanity Institute are actively working on the problem. However, it seems to me that this represents an appropriate allocation of resources, and that, for now at least, the issue is best addressed in a quiet academic research setting. It would be very difficult to justify anything on the scale of a government-funded “Manhattan Project” at this point in time. Nor does it seem wise to attempt to inject the issue into an already dysfunctional political process. Do we really want politicians with little or no understanding of the technology tweeting about the dangers of superintelligent machines? Given the very limited ability of the U.S. government in particular to accomplish almost anything at all, I also worry that hyping or politicizing a futuristic existential threat would be a distraction from the very real and immediate AI risks—including weaponization, security and bias—in which we really do need to begin investing significant resources in addressing right now.

迫切需要监管

A CRITICAL NEED FOR REGULATION

如果说我们在本章中探讨的风险有什么值得注意的地方，那就是随着人工智能的不断发展和普及，政府监管显然发挥着重要作用。然而，我认为过度监管或限制人工智能的一般研究是非常错误的。这样做在全球范围内可能无效，因为研究是在世界各地进行的。而且，正如我们所看到的，中国在推动人工智能前沿发展方面与美国和其他西方国家展开了激烈的竞争。对基础研究施加限制显然会让我们处于明显的劣势，我们绝对不能在寻求成为如此重大技术的前沿方面落后于中国。

If there’s one takeaway from the risks we’ve looked at in this chapter, it’s that there is clearly an important role for government regulation as AI continues to advance and become more ubiquitous. However, I think it would be very misguided to overly regulate or place limits on general research into artificial intelligence. Doing so would likely be ineffective on a global basis because the research is taking place all over the world. And, as we’ve seen, China in particular is engaged in intense competition with the United States and other Western countries in pushing the AI frontier forward. Placing restraints on basic research would clearly put us at a significant disadvantage, and we simply cannot afford to fall behind China in the quest to be at the leading edge of such a consequential technology.

相反，重点应该放在监管人工智能的具体应用上。在自动驾驶汽车或人工智能医疗诊断工具等领域，规则已经在制定中，因为这些应用与已经存在的监管框架相交叉。然而，我们需要更广泛的监督。人工智能最终将触及几乎所有领域，正如我们所见，面部识别或刑事司法系统中使用的算法等技术被用于做出非常高风险的决策，但几乎没有人能保证这些技术得到有效或公正的部署。

Rather, the focus should be on regulating specific applications of artificial intelligence. In areas such as self-driving cars or AI medical diagnostic tools, rules are already being developed because these applications intersect with a regulatory framework that is already in place. However, we need much broader oversight. Artificial intelligence will eventually touch virtually everything, and as we’ve seen, technologies like facial recognition or algorithms used in the criminal justice system are being used to make very high stakes decisions with virtually no guarantee that the technology is being deployed either effectively or justly.

鉴于人工智能的发展速度和所涉问题的复杂性，我认为期望美国国会或任何议会机构及时制定和颁布详细法规是不现实的。最好的做法可能是创建一个独立的政府机构，其监管权力专门针对人工智能的应用。这个机构大致相当于美国食品药品管理局、联邦航空管理局或美国证券交易委员会。在每种情况下，这些机构以及其他地方的同行，如欧洲药品管理局，都已经积累了深厚的内部专业知识，使他们能够解决其职权范围内的问题。人工智能领域也需要如此。国会将赋予人工智能监管机构广泛的权力和资金，但它将有权制定具体的法规，并且能够比立法机构更快、更有效地做到这一点。

Given the speed at which artificial intelligence is advancing and the complexity of the issues involved, I think it is unrealistic to expect the U.S. Congress, or indeed any parliamentary body, to write and enact detailed regulations in a timely fashion. The best course of action will probably be to create an independent governmental agency with regulatory powers specifically focused on applications of artificial intelligence. This would be an agency roughly comparable to the U.S. Food and Drug Administration, the Federal Aviation Administration or the Securities and Exchange Commission. In each case, these agencies—as well as their counterparts elsewhere, such as the European Medicines Agency—have developed deep in-house expertise that allows them to address the issues within their purview. The same needs to be true for the field of artificial intelligence. An AI regulatory agency would be given a broad mandate and allocated funds by Congress, but it would have the authority to write specific regulations and would be able to do this far more rapidly and effectively than the legislature.

那些具有自由主义倾向的人可能会反对，并正确地指出，这样的机构将遭受我们监管机构中已经存在的同样低效问题。人工智能监管机构肯定会与大型科技公司保持密切关系，我们可能会看到众所周知的“旋转门”，人们在行业和政府之间流动，而且存在监管俘获和技术行业不当影响的巨大风险。这些担忧是真实的，但尽管如此，我认为这样的机构显然是我们可用的最佳解决方案。如果替代方案只是什么也不做，那肯定会更糟糕。事实上，监管机构与开发和部署人工智能技术的公司之间的密切关系可能既是功能，也是缺陷。由于政府无法通过提供技术行业常见的那种薪水和股权补偿来争夺顶尖的人工智能人才，因此与私营部门的合作很可能是该机构能够跟上该领域最新发展的唯一途径。没有任何解决方案是完美的，但以拥有足够内部专业知识的监管机构为中心，建立产学研和政府之间的有效联盟，以保证事情朝着正确的方向发展，将在很大程度上确保人工智能的安全、包容和公正地部署。

Those with a libertarian orientation might well object and rightly point out that such an agency would suffer from the same inefficiencies already present in the rest of our regulatory apparatus. An AI regulatory agency would certainly have close relationships with large technology companies, we would likely see the proverbial “revolving-door,” in which people move between industry and government, and there would be a significant risk of regulatory capture and undue influence on the part of the technology industry. These concerns are real, but nonetheless I think that such an agency is pretty clearly the optimal solution available to us. If the alternative is to simply do nothing, that surely will be far worse. In fact, a close relationship between the regulatory agency and the companies developing and deploying AI technology is likely to be as much a feature as a bug. Because government cannot realistically compete for top AI talent by offering the kind of salaries and equity compensation common in the technology industry, cooperation with the private sector may well be the only way the agency will be able to keep pace with the latest developments in the field. No solution will be perfect, but a productive alliance between industry, academia and government, centered in a regulatory agency with sufficient in-house expertise to keep things moving in the right direction, would go a long way toward ensuring that AI is deployed safely, inclusively and justly.

^三十五

³⁵

结论

CONCLUSION

两种人工智能未来

TWO AI FUTURES

随着人工智能不断发展，其影响范围不断扩大到我们生活的方方面面，与该技术相关的风险将需要我们立即关注。2020 年，新冠病毒危机与广泛的社会动荡交织在一起，带来的发展表明，至少其中一些问题开始在公众讨论中占据突出地位。5 月，明尼阿波利斯警察杀害乔治·弗洛伊德引发全国性抗议活动，人们开始意识到面部识别技术中的种族偏见，亚马逊宣布暂停向执法机构销售其 Rekognition 系统一年，以便美国国会有时间考虑对该技术的监管。微软也宣布暂停销售，直至立法通过，而 IBM 则完全退出了面部识别市场^。1

AS ARTIFICIAL INTELLIGENCE CONTINUES TO ADVANCE AND EXTEND its reach into ever more facets of our lives, the risks associated with the technology will demand urgent attention. The intersection of the coronavirus crisis and widespread social upheaval in 2020 brought developments that suggest at least some of these issues are beginning to take a prominent place within public discourse. In the wake of the nationwide protests surrounding the killing of George Floyd by Minneapolis police officers in May, awareness of racial bias in facial recognition technology came to the forefront, and Amazon announced a one-year moratorium on sales of its Rekognition system to law enforcement agencies in order to give the U.S. Congress time to consider regulations on the technology. Microsoft announced a similar hiatus until legislation is passed, and IBM withdrew from the facial recognition market entirely.¹

冠状病毒大流行也带来了对非常规政策应对的新开放。由于经济停摆导致大量失业，国会得以迅速颁布几个月前可能夭折的政策。这些政策包括直接向纳税人发放 1,200 美元的刺激计划补助金、大幅（尽管是暂时的）增加失业保险金以及将该计划扩大到包括零工经济工作者。随着未来几年人工智能和机器人技术对就业市场的影响加速，所有这些想法现在都将被提上议程。事实上，已经有人呼吁在危机期间每月支付补助金（本质上是基本收入）^。2

The coronavirus pandemic has also brought about a new openness to unconventional policy responses. As the shutdown of the economy led to staggering job losses, Congress was able to rapidly enact policies that would have been dead on arrival just a few months earlier. These included $1,200 stimulus payments sent directly to taxpayers, a dramatic, albeit temporary, increase in unemployment insurance payments and an expansion of the program to include gig economy workers. All these ideas will now be on the table as the impact of artificial intelligence and robotics on the job market accelerates in the coming years. Indeed, there have already been calls for monthly payments—essentially a basic income—to be paid for the duration of the crisis.²

然而，随着人工智能的持续发展，必然会出现危险，因此，我们必须采取更加全面、更加一致的应对措施。这需要政府和私营部门之间进行有效协调，建立监管框架，并配备应对该领域快速发展所需的专业知识。所有这些都需要现在开始，因为我们可能已经落后了。

Still, a far more comprehensive and cohesive response to the dangers that will inevitably arrive with the continued rise of AI is critical. This will require effective coordination between government and the private sector and the creation of a regulatory framework coupled with the expertise necessary to respond to rapid advances in the field. And all this needs to begin now, as we are arguably already behind the curve.

尽管存在这些非常现实的担忧，但我坚信人工智能带来的好处将远远超过风险。事实上，考虑到我们在未来几十年将面临的挑战，我认为人工智能将是不可或缺的。我们需要人工智能将我们从技术瓶颈中带入一个广泛创新的新时代。

Despite these very real concerns, I firmly believe that the benefits from artificial intelligence will far outweigh the risks. Indeed, given the challenges we will face in the coming decades, I think AI will be indispensable. We will need artificial intelligence to launch us off our technological plateau into a new age of broad-based innovation.

气候变化是最明显的可预见威胁。2018 年，政府间气候变化专门委员会发布的一项分析表明，为了防止全球气温上升超过 1.5 摄氏度（希望这一阈值能够防止灾难性危害），我们需要在 2050 年之前将净碳排放量削减至零。为了有实际机会实现这一目标，我们需要在 2030 年之前将碳排放量减少约 45% ^。3

Climate change looms as the most clearly foreseeable threat. In 2018, the Intergovernmental Panel on Climate Change released an analysis indicating that in order to keep global temperatures from increasing by more than 1.5 degrees Celsius—a threshold that will hopefully prevent catastrophic harm—we will need to cut net carbon emissions to zero by the year 2050. And in order to have any realistic chance of achieving this, we need roughly a forty-five percent reduction by 2030.³

在新冠疫情爆发时，我们进行了大规模、前所未有的实验，这突显了这一挑战的严重性。正如比尔·盖茨在 2020 年 8 月的一篇博客文章中指出的那样，全球停摆——航空旅行几乎停止，全球的街道、高速公路和办公楼都空无一人——只导致排放量减少了约 8%。而这种暂时的减少是以数万亿美元的代价和几乎每个国家的失业率飙升为代价的。换句话说，假设我们能够在未来十年内通过主要依靠注重保护或行为改变（例如将我们的通勤转移到公共交通）的政策以某种方式将碳排放量减少近一半，这至少可以说是不现实的。正如盖茨所说，“我们不能仅仅通过减少飞行和驾驶来实现零排放，甚至不能主要通过减少飞行和驾驶来实现零排放。” ⁴

The magnitude of this challenge was brought into stark relief by the massive and unprecedented experiment we conducted as the coronavirus pandemic emerged. As Bill Gates pointed out in an August 2020 blog post, the global shutdown—in which air travel came to a near halt and streets, highways and office buildings emptied across the globe—led to only about an eight percent reduction in emissions. And that temporary decrease came at a cost of untold trillions of dollars and skyrocketing unemployment in nearly every country on earth. In other words, assuming that we can somehow cut carbon emissions nearly in half over the next decade by relying primarily on policies that focus on conservation or behavioral changes such as shifting our commute to public transit seems unrealistic to say the least. As Gates says, “We cannot get to zero emissions simply—or even mostly—by flying and driving less.”⁴

成功首先取决于创新。仅仅转向清洁、可再生的发电和驱动汽车的方式是不够的。发电厂和交通运输仅占全球排放量的 40% 左右。其余来自农业、制造业、建筑物排放和其他杂项来源。5^大幅减少全球排放需要在所有这些领域取得技术突破。再加上其他挑战，例如正在出现的全球淡水危机或不可避免的下一次大流行，很明显我们迫切需要全面创新。然而，正如我们在第 3 章中看到的那样，过去几十年来，新创意的产生速度实际上一直在放缓。正如研究美国创新的斯坦福大学和麻省理工学院的经济学家所写的那样，“无论我们在哪里，我们都会发现创意以及它们所暗示的指数增长越来越难找到。” ⁶

Success will depend first and foremost on innovation. And simply shifting to clean, renewable ways to generate electricity and power vehicles will not be sufficient. Electric power plants and transportation account for only about forty percent of global emissions. The remainder comes from agriculture, manufacturing, emissions from buildings and other miscellaneous sources.⁵ Dramatically reducing global emissions will require technological breakthroughs in all these areas. Add in other challenges, such as the emerging global fresh water crisis or the inevitable next pandemic, and it becomes clear that we are desperately in need of a burst of innovation across the board. Yet, as we saw in Chapter 3, the pace of new idea creation has actually been slowing over the past few decades. As the economists at Stanford and MIT who studied innovation in the United States wrote, “Everywhere we look we find that ideas, and the exponential growth they imply, are getting harder to find.”⁶

这种情况必须改变，而人工智能是实现这一目标的催化剂。面对这些挑战，没有什么比一种无处不在且价格合理的工具更重要，它可以极大地增强人类的智慧和创造力。关键目标是尽一切可能加速这种新资源的开发，同时发展我们的社会安全网和监管框架，使我们能够减轻伴随而来的风险，并确保人工智能的红利得到广泛和包容的分享。

This has to change, and artificial intelligence is the catalyst that can make it happen. In the face of these challenges, nothing could be more consequential than a ubiquitous and affordable utility that dramatically amplifies human intellect and creativity. The key objective is to do everything possible to accelerate the development of this new resource while at the same time evolving our social safety net and regulatory framework in ways that will allow us to mitigate the accompanying risks and ensure that the dividends from AI are shared widely and inclusively.

在我们沿着这条道路前进的过程中，我认为，我们所构建的未来最终可能会落在两个虚构极端之间的某个范围内。最乐观的情景来自电视剧《星际迷航》。在这个后稀缺世界里，先进的技术创造了物质上的丰富，消除了贫困，解决了环境问题，治愈了大多数疾病。没有人需要辛苦地从事一份没有回报的工作来维持生存。这个世界上的人们受过高等教育，追求他们认为有意义的挑战。没有传统工作并没有导致懒惰或缺乏意义和尊严。在《星际迷航》的世界中，人们的价值取决于他们内在的人性——而不是主要基于他们的经济产出。虽然《星际迷航》中描绘的许多技术很可能是无法实现的，或者至少是在遥远的未来，但我认为该剧合理地描绘了一个未来，在那里，先进的技术带来了广泛的繁荣，解决了人类在地球上的挑战，并使我们能够触及星空。

As we navigate this path forward, I think that the future we build may ultimately fall somewhere on a spectrum bounded by two fictional extremes. The most optimistic scenario comes from the television show Star Trek. In this post-scarcity world, advanced technology has created material abundance, eliminated poverty, addressed environmental concerns and cured most disease. No one needs to toil in drudgery at an unrewarding job simply to sustain his or her survival. People in this world are highly educated and pursue challenges that they find rewarding. The absence of a traditional job has not led to idleness or a lack of meaning and dignity. In the Star Trek universe, people are valued for their intrinsic humanity—not primarily based on their economic output. Though many of the technologies portrayed in Star Trek are likely unrealizable or, at a minimum, lie in the distant future, I think the show offers a reasonable rendition of a future in which advanced technology leads to broad-based prosperity, solves humanity’s terrestrial challenges and allows us to reach for the stars.

另一个更加反乌托邦的未来可能更接近于《黑客帝国》。我担心的不是人工智能会以某种方式奴役我们，而是现实世界可能变得如此不平等，大多数普通人缺乏提升自己前景的机会，以至于很大一部分人会选择逃避现实。随着人工智能和虚拟现实技术在未来几年和几十年的加速发展，它们可能会结合起来创造出极具吸引力和逼真的模拟环境，对许多人来说，这些模拟环境可能比他们实际生活的世界要优越得多。事实上，一组经济学家在 2017 年的一项分析发现，越来越多脱离劳动力市场的年轻人将大量时间花在玩电子游戏上。7^这项技术很快就会让这些虚拟环境变得如此令人上瘾，以至于它们可以被合理地视为一种毒品。

The alternate, and far more dystopian, future might be something closer to The Matrix. My fear is not that artificial intelligence will somehow enslave us, but rather that the real world might become so unequal, and so lacking in opportunity for most typical people to advance their prospects, that a large fraction of the population will choose to escape into alternative realities. As both AI and virtual reality technology accelerate over the coming years and decades, they will likely combine to create extraordinarily compelling and realistic simulations that, to many people, may seem far superior to the world in which they actually live. Indeed, a 2017 analysis by a group of economists found that the increasing number of young men who are detached from labor markets are spending an outsized portion of their time playing video games.⁷ The technology will soon arrive to make these virtual environments so addictive that they might reasonably be viewed as a kind of drug.

如果人工智能和机器人颠覆了就业市场，就业机会消失或质量下降，政府最终很可能会被迫向公民提供某种形式的支持（可能是基本收入），以维持社会秩序。但是，如果他们忽视确保民众继续优先考虑教育并保持目标感，结果可能是普遍的冷漠和脱离。我们可能会走向一个分裂为少数精英的社会，他们仍然扎根于现实世界，而大众则越来越多地逃避技术幻想，或者被犯罪或其他形式的成瘾所吸引。最终，我们的人口受教育程度会降低，民主制度将远没有那么包容和有效，创新步伐也会更慢，因为许多最聪明的人被引诱进入一个越来越有吸引力的虚拟世界，可能不再有强烈的动力去争取现实世界的成功。在这种情况下，经济和社会逆风将使我们更难克服我们面临的全球挑战。

If artificial intelligence and robotics upend the job market and employment opportunities evaporate or decline in quality, governments will in all likelihood eventually be forced to provide some form of support—perhaps a basic income—to citizens in order to maintain social order. If they neglect, however, to also ensure that the population continues to prioritize education and maintain a sense of purpose, the result is likely to be widespread apathy and disengagement. We might trend toward a society that balkanizes into a small elite that remains anchored in the real world, while the masses increasingly escape into a technological fantasy or perhaps are drawn to crime or other forms of addiction. We would then end up with a less educated population, a far less inclusive and effective democracy and a slower pace of innovation because many of the brightest individuals, lured into an ever more compelling virtual realm, might no longer see a strong incentive to strive for real-world success. In this scenario, economic and social headwinds will make it far more difficult for us to overcome the global challenges we face.

我想几乎每个人都会同意，我们应该努力创造一个更接近 《星际迷航》的未来。然而，这不会自动发生。我们需要制定明确的政策，旨在改变我们的轨迹，走向那个目的地。很有可能，我们还需要很长时间才能到达目的地，但如果我们能从解决收入分配问题开始，同时保持强大的激励，让人们接受教育并追求有意义的挑战，我们就会朝着正确的方向前进。

I think nearly everyone would agree that we should strive for a future that comes closer to Star Trek. This, however, will not happen by default. We will need to craft explicit policies designed to shift our trajectory toward that destination. In all likelihood, it will be a very long time before we arrive, but if we can begin by solving the problem of income distribution while maintaining a strong incentive for people to educate themselves and pursue meaningful challenges, we will be headed in the right direction.

⁷

致谢

ACKNOWLEDGMENTS

过去几年，很多人通过对话和技术演示帮助我理解了人工智能。我特别感谢参与《智能建筑师》对话的 23 位杰出研究人员和企业家。他们确实是人工智能领域最聪明的人，他们的见解和预测为本书的大部分内容提供了参考。

A great many people contributed to my understanding of artificial intelligence through conversations and technology demonstrations over the past few years. I’m especially grateful to the twenty-three prominent researchers and entrepreneurs who participated in the conversations recorded in Architects of Intelligence. They are truly among the brightest minds in the field of AI, and their insights and predictions informed much of the material in this book.

我的编辑，美国的 TJ Kelleher 和英国的 Sarah Caro 帮助我完善论点，并将手稿结构化为最佳形式。我的经纪人 Don Fehr 再次在 Basic Books 为这个项目找到了合适的归宿。

My editors, TJ Kelleher in the United States and Sarah Caro in the United Kingdom, were instrumental in helping me to refine my arguments and structure the manuscript into its optimal form. My agent, Don Fehr, once again found the proper home for this project at Basic Books.

我写这本书的大约八个月时间恰逢新冠疫情爆发和随之而来的停工。在此期间，我非常幸运能够安全地待在家里，专心写作，我非常感谢所有没有这种奢侈的医护人员和一线工作人员。

The roughly eight months that I spent writing this book coincided with the emergence of the coronavirus pandemic and the ensuing shutdown. I was very fortunate to be able to remain safely at home and focus on my writing during this period, and I am profoundly grateful to all the health care professionals and frontline workers who had no such luxury.

最后，我要感谢我的妻子晓晓和女儿Elaine，感谢她们在我全身心投入这个项目时给予我的鼓励和支持。

Finally, I thank my wife, Xiaoxiao, and my daughter, Elaine, for their encouragement and support as I immersed myself in this project.

发现你的下一本精彩读物

Discover Your Next Great Read

获取有关您最喜欢的作家的预览、书籍推荐和新闻。

Get sneak peeks, book recommendations, and news about your favorite authors.

点击此处了解更多。

基础图书徽标

笔记

NOTES

第一章新兴颠覆

CHAPTER 1. THE EMERGING DISRUPTION

1. Ewen Callaway，“‘它将改变一切’：DeepMind 的 AI 在解决蛋白质结构方面取得了巨大飞跃”， 《自然》 ，2020 年 11 月 30 日，www.nature.com/articles/ d41586-020-03348-4 。

1. Ewen Callaway, “‘It will change everything’: DeepMind’s AI makes gigantic leap in solving protein structures,” Nature, November 30, 2020, www.nature.com/articles/d41586-020-03348-4.

2. Andrew Senior、Demis Hassabis、John Jumper 和 Pushmeet Kohli，《AlphaFold：使用 AI 进行科学发现》，DeepMind 研究博客，2020 年 1 月 15 日，deepmind.com/blog/article/AlphaFold-Using-AI-for-scientific-discovery。

2. Andrew Senior, Demis Hassabis, John Jumper and Pushmeet Kohli, “AlphaFold: Using AI for scientific discovery,” DeepMind Research Blog, January 15, 2020, deepmind.com/blog/article/AlphaFold-Using-AI-for-scientific-discovery.

3. Ian Sample，“谷歌 DeepMind 预测蛋白质的 3D 形状”， 《卫报》 ，2018 年 12 月 2 日，www.theguardian.com/science/2018 / dec/02/google-deepminds-ai-program-alphafold-predicts-3d-shapes-of-proteins。

3. Ian Sample, “Google’s DeepMind predicts 3D shapes of proteins,” The Guardian, December 2, 2018, www.theguardian.com/science/2018/dec/02/google-deepminds-ai-program-alphafold-predicts-3d-shapes-of-proteins.

4. Lyxor Robotics 和AI UCITS ETF，股票市场代码为 ROAI。

4. Lyxor Robotics and AI UCITS ETF, stock market ticker ROAI.

5.例如，请参阅：Carl Benedikt Frey 和 Michael Osborne，《就业的未来：工作对计算机化有多敏感？》，牛津大学牛津马丁学院，工作论文，2013 年 9 月 17 日，www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf，第 38 页。

5. See, for example: Carl Benedikt Frey and Michael Osborne, “The future of employment: How susceptible are jobs to computerisation?,” Oxford Martin School, University of Oxford, Working Paper, September 17, 2013, www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf, p. 38.

6. Matt McFarland，“伊隆·马斯克：‘我们利用人工智能召唤恶魔’” ， 《华盛顿邮报》 ，2014 年 10 月 24 日，www.washingtonpost.com/news/innovations/wp/2014 / 10/24/elon-musk-with-artificial-intelligence-we-are-summoning-the-demon /。

6. Matt McFarland, “Elon Musk: ‘With artificial intelligence we are summoning the demon,’” Washington Post, October 24, 2014, www.washingtonpost.com/news/innovations/wp/2014/10/24/elon-musk-with-artificial-intelligence-we-are-summoning-the-demon/.

7. Anand S. Rao 和 Gerard Verweij，“确定奖项规模：人工智能对您的业务的真正价值是什么？您如何利用它？”，普华永道，2018 年 10 月，www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf。

7. Anand S. Rao and Gerard Verweij, “Sizing the prize: What’s the real value of AI for your business and how can you capitalise?,” PwC, October 2018, www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf.

第 2 章人工智能作为新电力

CHAPTER 2. AI AS THE NEW ELECTRICITY

1. “神经形态计算”，英特尔公司，2020 年 5 月 3 日访问，www.intel.com/content/www/us/en/research/neuromorphic-computing.html。

1. “Neuromorphic computing,” Intel Corporation, accessed May 3, 2020, www.intel.com/content/www/us/en/research/neuromorphic-computing.html.

2. Sara Castellanos，“英特尔将发布神经形态计算系统”， 《华尔街日报》 ，2020 年 3 月 18 日，www.wsj.com/articles/intel-to -release-neuromorphic -computing-system-11584540000。

2. Sara Castellanos, “Intel to release neuromorphic-computing system,” Wall Street Journal, March 18, 2020, www.wsj.com/articles/intel-to-release-neuromorphic-computing-system-11584540000.

3. Linda Hardesty，“维基解密公布亚马逊数据中心位置”，SDXCentral，2018 年 10 月 12 日，www.sdxcentral.com/articles/news/wikileaks-publishes-the-location-of-amazons-data-centers/2018/10/。

3. Linda Hardesty, “WikiLeaks publishes the location of Amazon’s data centers,” SDXCentral, October 12, 2018, www.sdxcentral.com/articles/news/wikileaks-publishes-the-location-of-amazons-data-centers/2018/10/.

4. “Flexera 发布的 RightScale 2019 云现状报告”，Flexera，2019 年，resources.flexera.com/web/media/documents/rightscale-2019-state-of-the-cloud-report-from-flexera.pdf，第 2 页。

4. “RightScale 2019 State of the Cloud Report from Flexera,” Flexera, 2019, resources.flexera.com/web/media/documents/rightscale-2019-state-of-the-cloud-report-from-flexera.pdf, p. 2.

5. Pierr Johnson，“随着亚马逊、微软和谷歌等公有云的出现，大数据已成为众所周知的大事” ， 《福布斯》 ，2017 年 6 月 15 日，www.forbes.com/sites/johnsonpierr/2017 / 06/15/with-the-public-clouds-of-amazon-microsoft-and-google-big-data-is-the-proverbial-big-deal /。

5. Pierr Johnson, “With the public clouds of Amazon, Microsoft and Google, big data is the proverbial big deal,” Forbes, June 15, 2017, www.forbes.com/sites/johnsonpierr/2017/06/15/with-the-public-clouds-of-amazon-microsoft-and-google-big-data-is-the-proverbial-big-deal/.

6. Richard Evans 和 Jim Gao，“DeepMind AI 将 Google 数据中心的制冷费用降低了 40%”，DeepMind 研究博客，2016 年 7 月 20 日，deepmind.com/blog/article/deepmind-ai-reduces-google-data-centre-cooling-bill-40。

6. Richard Evans and Jim Gao, “DeepMind AI reduces Google data centre cooling bill by 40%,” DeepMind Research Blog, July 20, 2016, deepmind.com/blog/article/deepmind-ai-reduces-google-data-centre-cooling-bill-40.

7. Urs Hölzle，“数据中心比以往任何时候都更节能”，Google 博客，2020 年 2 月 27 日，www.blog.google/outreach-initiatives/sustainability/data-centers-energy-efficient/。

7. Urs Hölzle, “Data centers are more energy efficient than ever,” Google Blog, February 27, 2020, www.blog.google/outreach-initiatives/sustainability/data-centers-energy-efficient/.

8. Ron Miller，“AWS 收入增长略有下滑，但仍是亚马逊的摇钱树”， TechCrunch ，2019 年 7 月 25 日，techcrunch.com/2019/07/25/aws- revenue -growth-slips-a-bit-but-remains-amazons-golden -goose/。

8. Ron Miller, “AWS revenue growth slips a bit, but remains Amazon’s golden goose,” TechCrunch, July 25, 2019, techcrunch.com/2019/07/25/aws-revenue-growth-slips-a-bit-but-remains-amazons-golden-goose/.

9. John Bonazzo，“谷歌在员工抗议后退出五角大楼‘JEDI’项目”， 《观察家报》，2018 年 10 月 10 日，observer.com/2018/10/google-pentagon-jedi/ 。

9. John Bonazzo, “Google exits Pentagon ‘JEDI’ project after employee protests,” Observer, October 10, 2018, observer.com/2018/10/google-pentagon-jedi/.

10. Annie Palmer，“亚马逊起诉后，法官暂时阻止微软五角大楼云计算合同”，CNBC，2020 年 2 月 13 日，www.cnbc.com/2020/02/13/ amazon -gets-restraining-order-to-block-microsoft-work-on-pentagon-jedi.html。

10. Annie Palmer, “Judge temporarily blocks Microsoft Pentagon cloud contract after Amazon suit,” CNBC, February 13, 2020, www.cnbc.com/2020/02/13/amazon-gets-restraining-order-to-block-microsoft-work-on-pentagon-jedi.html.

11. Lauren Feiner，“国防部要求法官让其重新考虑将 100 亿美元合同授予微软而非亚马逊的决定”，CNBC，2020 年 3 月 13 日，www.cnbc.com/2020/03/13/pentagon-asks-judge-to-let-it-reconsider-its-jedi-cloud-contract-award.html。

11. Lauren Feiner, “DoD asks judge to let it reconsider decision to give Microsoft $10 billion contract over Amazon,” CNBC, March 13, 2020, www.cnbc.com/2020/03/13/pentagon-asks-judge-to-let-it-reconsider-its-jedi-cloud-contract-award.html.

12. “AWS 上的 TensorFlow”，Amazon Web Services，2020 年 5 月 4 日访问，aws.amazon.com/tensorflow/。

12. “TensorFlow on AWS,” Amazon Web Services, accessed May 4, 2020, aws.amazon.com/tensorflow/.

13. Kyle Wiggers，“英特尔推出 Pohoiki Springs，一款适用于 AI 工作负载的强大神经形态研究系统”， VentureBeat ，2020 年 3 月 18 日，venturebeat.com/2020/03/18/intel-debuts- pohoiki -springs-a-powerful-neuromorphic-research-system-for-ai-workloads /。

13. Kyle Wiggers, “Intel debuts Pohoiki Springs, a powerful neuromorphic research system for AI workloads,” VentureBeat, March 18, 2020, venturebeat.com/2020/03/18/intel-debuts-pohoiki-springs-a-powerful-neuromorphic-research-system-for-ai-workloads/.

14. Jeremy Kahn，《探究大型科技公司对人类级人工智能的追求》， 《财富》，2020 年 1 月 20 日，fortune.com/longform/ai-artificial-intelligence-big-tech-microsoft-alphabet-openai/ 。

14. Jeremy Kahn, “Inside big tech’s quest for human-level A.I.,” Fortune, January 20, 2020, fortune.com/longform/ai-artificial-intelligence-big-tech-microsoft-alphabet-openai/.

15. Martin Ford，对李飞飞的采访，载于《智能建筑师：来自人工智能建造者的真相》，Packt Publishing，2018 年，第 150 页。

15. Martin Ford, Interview with Fei-Fei Li, in Architects of Intelligence: The Truth about AI from the People Building It, Packt Publishing, 2018, p. 150.

16. “AWS 上的深度学习”，Amazon Web Services，2020 年 5 月 4 日访问，aws.amazon.com/deep-learning/。

16. “Deep Learning on AWS,” Amazon Web Services, accessed May 4, 2020, aws.amazon.com/deep-learning/.

17. Kyle Wiggers，“麻省理工学院研究人员：亚马逊的 Rekognition 显示出性别和种族偏见”， VentureBeat，2019 年 1 月 24 日，venturebeat.com/2019/01/24/amazon-rekognition-bias-mit/ 。

17. Kyle Wiggers, “MIT researchers: Amazon’s Rekognition shows gender and ethnic bias,” VentureBeat, January 24, 2019, venturebeat.com/2019/01/24/amazon-rekognition-bias-mit/.

18. “新方案教会大众构建人工智能”， 《经济学人》 ，2018 年 10 月 27 日，www.economist.com/business/2018/10/27/new-schemes-teach-the-masses-to-build - ai。

18. “New schemes teach the masses to build AI,” The Economist, October 27, 2018, www.economist.com/business/2018/10/27/new-schemes-teach-the-masses-to-build-ai.

19. Chris Hoffman，“什么是 5G，它的速度有多快？”， How-to Geek ，2020 年 1 月 3 日，www.howtogeek.com/340002/what-is-5g-and-how- fast -will-it-be/。

19. Chris Hoffman, “What is 5G, and how fast will it be?,” How-to Geek, January 3, 2020, www.howtogeek.com/340002/what-is-5g-and-how-fast-will-it-be/.

第 3 章超越炒作

CHAPTER 3. BEYOND HYPE

1.特斯拉，“特斯拉自动驾驶日（视频）”，YouTube，2019 年 4 月 22 日，www.youtube.com/watch?reload=9&v=Ucp0TTmvqOE。

1. Tesla, “Tesla Autonomy Day (video),” YouTube, April 22, 2019, www.youtube.com/watch?reload=9&v=Ucp0TTmvqOE.

2. Sean Szymkowski，“特斯拉全自动驾驶模式受到 NHTSA 的密切关注”， Road Show，2020 年 10 月 22 日，www.cnet.com/ roadshow/news/teslas-full - self-driving-mode-nhtsa/。

2. Sean Szymkowski, “Tesla’s full self-driving mode under the watchful eye of NHTSA,” Road Show, October 22, 2020, www.cnet.com/roadshow/news/teslas-full-self-driving-mode-nhtsa/.

3. Rob Csongor，“特斯拉提高自动驾驶汽车制造商的标准”，NVIDIA 博客，2019 年 4 月 23 日，blogs.nvidia.com/blog/2019/04/23/tesla-self-driving/。

3. Rob Csongor, “Tesla raises the bar for self-driving carmakers,” NVIDIA Blog, April 23, 2019, blogs.nvidia.com/blog/2019/04/23/tesla-self-driving/.

4. Jeffrey Van Camp，“我的 Jibo 快死了，这让我心碎”， 《连线》，2019 年 3 月 9 日，www.wired.com/story/jibo-is-dying-eulogy/。

4. Jeffrey Van Camp, “My Jibo is dying and it’s breaking my heart,” Wired, March 9, 2019, www.wired.com/story/jibo-is-dying-eulogy/.

5. Mark Gurman 和 Brad Stone，“据称亚马逊正在开展另一项大赌注：家用机器人” ， 彭博社，2018 年 4 月 23 日，www.bloomberg.com/news/articles/2018-04-23/amazon-is-said-to-be - working-on-another-big-bet-home-robots 。

5. Mark Gurman and Brad Stone, “Amazon is said to be working on another big bet: Home robots,” Bloomberg, April 23, 2018, www.bloomberg.com/news/articles/2018-04-23/amazon-is-said-to-be-working-on-another-big-bet-home-robots.

6.马丁·福特，采访罗德尼·布鲁克斯， 《智能建筑师：来自人工智能建造者的真相》，Packt Publishing，2018 年，第 432 页。

6. Martin Ford, Interview with Rodney Brooks, in Architects of Intelligence: The Truth about AI from the People Building It, Packt Publishing, 2018, p. 432.

7. “用机械手解决魔方”，OpenAI，2019 年 10 月 15 日，openai.com/blog/solving-rubiks-cube/。（包括视频。）

7. “Solving Rubik’s Cube with a robot hand,” OpenAI, October 15, 2019, openai.com/blog/solving-rubiks-cube/. (Includes videos.)

8. Will Knight，“为什么解开魔方并不代表机器人至上”， 《连线》，2019 年 10 月 16 日，www.wired.com/ story/why-solving-rubiks-cube - not-signal-robot-supremacy/。

8. Will Knight, “Why solving a Rubik’s Cube does not signal robot supremacy,” Wired, October 16, 2019, www.wired.com/story/why-solving-rubiks-cube-not-signal-robot-supremacy/.

9. Noam Scheiber，“在亚马逊仓库内，机器人的工作方式对人类产生影响”， 《纽约时报》 ，2019 年 7 月 3 日，www.nytimes.com/2019/07 /03 /business/economy/amazon-warehouse-labor-robots.html。

9. Noam Scheiber, “Inside an Amazon warehouse, robots’ ways rub off on humans,” New York Times, July 3, 2019, www.nytimes.com/2019/07/03/business/economy/amazon-warehouse-labor-robots.html.

10. Eugene Kim，“亚马逊斥资 7.75 亿美元收购机器人公司 Kiva 的交易开始显得非常明智”， 《商业内幕》，2016 年 6 月 15 日，www.businessinsider.com/kiva-robots-save-money-for-amazon-2016-6 。

10. Eugene Kim, “Amazon’s $775 million deal for robotics company Kiva is starting to look really smart,” Business Insider, June 15, 2016, www.businessinsider.com/kiva-robots-save-money-for-amazon-2016-6.

11 . Will Evans，“亚马逊无情的配额正在伤害员工”， 《大西洋月刊》，2019 年 11 月 25 日，www.theatlantic.com/technology/archive/2019/11/amazon-warehouse-reports-show-worker-injuries/602530/。

11. Will Evans, “Ruthless quotas at Amazon are maiming employees,” The Atlantic, November 25, 2019, www.theatlantic.com/technology/archive/2019/11/amazon-warehouse-reports-show-worker-injuries/602530/.

12. Jason Del Ray，“机器人如何改变亚马逊仓库工作——有好有坏”， Recode，2019 年 12 月 11 日，www.vox.com/ recode/2019/12/11/20982652/ robots -amazon-warehouse-jobs-automation。

12. Jason Del Ray, “How robots are transforming Amazon warehouse jobs—for better and worse,” Recode, December 11, 2019, www.vox.com/recode/2019/12/11/20982652/robots-amazon-warehouse-jobs-automation.

13. Michael Sainato，“‘我不是机器人’：亚马逊员工谴责仓库不安全、艰苦的工作条件”， 《卫报》 ，2020 年 2 月 5 日，www.theguardian.com/technology/2020 / feb/05/amazon-workers-protest-unsafe-grueling-conditions-warehouse 。

13. Michael Sainato, “‘I’m not a robot’: Amazon workers condemn unsafe, grueling conditions at warehouse,” The Guardian, February 5, 2020, www.theguardian.com/technology/2020/feb/05/amazon-workers-protest-unsafe-grueling-conditions-warehouse.

14 . Jeffrey Dastin，“独家：亚马逊推出可打包订单并取代工作的机器”，路透社，2019 年 5 月 13 日，www.reuters.com/article/us-amazon-com-automation-exclusive/exclusive-amazon-rolls-out-machines-that-pack-orders-and-replace-jobs-idUSKCN1SJ0X1。

14. Jeffrey Dastin, “Exclusive: Amazon rolls out machines that pack orders and replace jobs,” Reuters, May 13, 2019, www.reuters.com/article/us-amazon-com-automation-exclusive/exclusive-amazon-rolls-out-machines-that-pack-orders-and-replace-jobs-idUSKCN1SJ0X1.

15. Matt Simon，“亚马逊仓库内部，人机合一”， 《连线》，2019 年 6 月 5 日，www.wired.com/story/amazon-warehouse-robots/ 。

15. Matt Simon, “Inside the Amazon warehouse where humans and machines become one,” Wired, June 5, 2019, www.wired.com/story/amazon-warehouse-robots/.

16. James Vincent，“亚马逊最新的机器人冠军使用深度学习来整理货架”， The Verge，2016 年 7 月 5 日，www.theverge.com/2016/7/5/12095788/amazon-picking-robot-challenge-2016 。

16. James Vincent, “Amazon’s latest robot champion uses deep learning to stock shelves,” The Verge, July 5, 2016, www.theverge.com/2016/7/5/12095788/amazon-picking-robot-challenge-2016.

17. Jeffrey Dastin，“亚马逊的贝索斯表示，机械手将在未来 10 年内投入商业使用”，路透社，2019 年 6 月 6 日，www.reuters.com/article/us-amazon-com-conference/amazons-bezos-says-robotic-hands-will-be-ready-for-commercial-use-in-next-10-years-idUSKCN1T72JB。

17. Jeffrey Dastin, “Amazon’s Bezos says robotic hands will be ready for commercial use in next 10 years,” Reuters, June 6, 2019, www.reuters.com/article/us-amazon-com-conference/amazons-bezos-says-robotic-hands-will-be-ready-for-commercial-use-in-next-10-years-idUSKCN1T72JB.

18. Tech Insider，“数千个机器人在仓库内打包杂货（视频）”，YouTube，2018 年 5 月 9 日，www.youtube.com/watch?reload=9&v=4DKrcpa8Z_E。

18. Tech Insider, “Inside a warehouse where thousands of robots pack groceries (video),” YouTube, May 9, 2018, www.youtube.com/watch?reload=9&v=4DKrcpa8Z_E.

19. James Vincent，“欢迎来到未来的自动化仓库”， The Verge，2018 年 5 月 8 日，www.theverge.com/2018/5/8/17331250/automated-warehouses-jobs-ocado-andover-amazon 。

19. James Vincent, “Welcome to the automated warehouse of the future,” The Verge, May 8, 2018, www.theverge.com/2018/5/8/17331250/automated-warehouses-jobs-ocado-andover-amazon.

20.同上。

20. Ibid.

21. “ABB 与 Covariant 合作部署集成 AI 机器人解决方案”，ABB 新闻稿，2020 年 2 月 25 日，new.abb.com/news/detail/57457/abb-and-covariant-partner-to-deploy-integrated-ai-robotic-solutions。

21. “ABB and Covariant partner to deploy integrated AI robotic solutions,” ABB Press Release, February 25, 2020, new.abb.com/news/detail/57457/abb-and-covariant-partner-to-deploy-integrated-ai-robotic-solutions.

22. Evan Ackerman，“Covariant 使用简单的机器人和巨大的神经网络实现仓库拣选自动化”， IEEE Spectrum ，2020 年 1 月 29 日，spectrum.ieee.org/ automaton /robotics/industrial-robots/covariant-ai-gigantic-neural-network-to-automate-warehouse-picking 。

22. Evan Ackerman, “Covariant uses simple robot and gigantic neural net to automate warehouse picking,” IEEE Spectrum, January 29, 2020, spectrum.ieee.org/automaton/robotics/industrial-robots/covariant-ai-gigantic-neural-network-to-automate-warehouse-picking.

23. Jonathan Vanian，“工业机器人巨头与新兴 AI 初创公司合作”， 《财富》，2020 年 2 月 25 日，fortune.com/2020/02/25/industrial-robotics-ai-covariant/ 。

23. Jonathan Vanian, “Industrial robotics giant teams up with a rising A.I. startup,” Fortune, February 25, 2020, fortune.com/2020/02/25/industrial-robotics-ai-covariant/.

24. Alexander Lavin、J. Swaroop Guntupalli、Miguel Lázaro-Gredilla 等人，“使用递归皮层网络解释视觉皮层现象”，Vicarious 研究论文，2018 年 7 月 30 日，www.biorxiv.org/content/biorxiv/early/2018/07/30/380048.full.pdf。

24. Alexander Lavin, J. Swaroop Guntupalli, Miguel Lázaro-Gredilla, et al., “Explaining visual cortex phenomena using recursive cortical network,” Vicarious Research Paper, July 30, 2018, www.biorxiv.org/content/biorxiv/early/2018/07/30/380048.full.pdf.

25. Tom Simonite，“这些工业机器人在执行每项任务时都变得更加熟练”， 《连线》，2020 年 3 月 10 日，www.wired.com/ story/these-industrial-robots - adept-every-task/。

25. Tom Simonite, “These industrial robots get more adept with every task,” Wired, March 10, 2020, www.wired.com/story/these-industrial-robots-adept-every-task/.

26. Adam Satariano 和 Cade Metz，“仓库机器人学会整理棘手的东西”， 《纽约时报》 ，2020 年 1 月 29 日，www.nytimes.com/2020/01/29/ technology /warehouse-robot.html。

26. Adam Satariano and Cade Metz, “A warehouse robot learns to sort out the tricky stuff,” New York Times, January 29, 2020, www.nytimes.com/2020/01/29/technology/warehouse-robot.html.

27. Matthew Boyle，“第二通道的机器人：超市生存意味着要与亚马逊匹敌”， 彭博社，2019 年 12 月 3 日，www.bloomberg.com/features/2019-automated-grocery-stores/ 。

27. Matthew Boyle, “Robots in aisle two: Supermarket survival means matching Amazon,” Bloomberg, December 3, 2019, www.bloomberg.com/features/2019-automated-grocery-stores/.

28.同上。

28. Ibid.

29. Nathaniel Meyersohn，“杂货店在冠状病毒爆发期间转向机器人”，CNN Business，2020 年 4 月 7 日，www.cnn.com/2020/04/07/business/grocery-stores-robots-automation/index.html。

29. Nathaniel Meyersohn, “Grocery stores turn to robots during the coronavirus,” CNN Business, April 7, 2020, www.cnn.com/2020/04/07/business/grocery-stores-robots-automation/index.html.

30. Shoshy Ciment，“沃尔玛将在全国范围内加大门店自动化程度，将机器人引入另外 650 家门店”， 《商业内幕》，2020 年 1 月 13 日，www.businessinsider.com/walmart-adding-robots - help-stock-shelves-to-650-more-stores -2020-1。

30. Shoshy Ciment, “Walmart is bringing robots to 650 more stores as the retailer ramps up automation in stores nationwide,” Business Insider, January 13, 2020, www.businessinsider.com/walmart-adding-robots-help-stock-shelves-to-650-more-stores-2020-1.

31. Jennifer Smith，“微型配送中心助力食品杂货配送小型化”， 《华尔街日报》，2020 年 1 月 27 日，www.wsj.com/articles/grocery-delivery-goes-small-with-micro-fulfillment-centers-11580121002 。

31. Jennifer Smith, “Grocery delivery goes small with micro-fulfillment centers,” Wall Street Journal, January 27, 2020, www.wsj.com/articles/grocery-delivery-goes-small-with-micro-fulfillment-centers-11580121002.

32. Nick Wingfield，“走进 Amazon Go，未来商店”， 纽约时报，2018 年 1 月 21 日，www.nytimes.com/2018/01/21/technology/inside-amazon-go-a-store-of-the-future.html 。

32. Nick Wingfield, “Inside Amazon Go, a store of the future,” New York Times, January 21, 2018, www.nytimes.com/2018/01/21/technology/inside-amazon-go-a-store-of-the-future.html.

33 . Spencer Soper，“亚马逊将考虑在 2021 年前开设多达 3,000 家无人收银员商店”， 彭博社，2018 年 9 月 29 日，www.bloomberg.com/news/articles/2018-09-19/amazon-is-said-to-plan-up-to-3-000-cashierless-stores-by-2021。

33. Spencer Soper, “Amazon will consider opening up to 3,000 cashierless stores by 2021,” Bloomberg, September 29, 2018, www.bloomberg.com/news/articles/2018-09-19/amazon-is-said-to-plan-up-to-3-000-cashierless-stores-by-2021.

34. Paul Sawyers，“软银领投 3000 万美元投资 Accel Robotics，打造人工智能无人收银商店”， VentureBeat ，2019 年 12 月 3 日，venturebeat.com/2019 / 12/03/softbank-leads-30-million-investment-in-accel-robotics-for-ai-enabled-cashierless-stores /。

34. Paul Sawyers, “SoftBank leads $30 million investment in Accel Robotics for AI-enabled cashierless stores,” VentureBeat, December 3, 2019, venturebeat.com/2019/12/03/softbank-leads-30-million-investment-in-accel-robotics-for-ai-enabled-cashierless-stores/.

35. Jurica Dujmovic，“随着新冠疫情的严重打击，亚马逊开始向零售商授权无人收银技术”，MarketWatch，2020 年 3 月 31 日，www.marketwatch.com/story/as-coronavirus-hits-hard-amazon-starts-licensing-cashier-free-technology-to-retailers-2020-03-31。

35. Jurica Dujmovic, “As coronavirus hits hard, Amazon starts licensing cashier-free technology to retailers,” MarketWatch, March 31, 2020, www.marketwatch.com/story/as-coronavirus-hits-hard-amazon-starts-licensing-cashier-free-technology-to-retailers-2020-03-31.

36 . Eric Rosenbaum，“随着快餐业员工流失危机加剧，Panera 每年几乎失去 100% 的员工”，CNBC，2019 年 8 月 29 日，www.cnbc.com /2019/08/29/fast-food-restaurants-in-america-are-losing-100percent-of-workers-every-year.html。

36. Eric Rosenbaum, “Panera is losing nearly 100% of its workers every year as fast-food turnover crisis worsens,” CNBC, August 29, 2019, www.cnbc.com/2019/08/29/fast-food-restaurants-in-america-are-losing-100percent-of-workers-every-year.html.

37.同上。

37. Ibid.

38. Kate Krader，“全球首款机器人汉堡即将登陆湾区”， 彭博社，2018 年 6 月 21 日，www.bloomberg.com/news/features/2018-06-21/the-world-s - first-robotic-burger-is-ready-to-hit-the-bay-area 。

38. Kate Krader, “The world’s first robot-made burger is about to hit the Bay Area,” Bloomberg, June 21, 2018, www.bloomberg.com/news/features/2018-06-21/the-world-s-first-robotic-burger-is-ready-to-hit-the-bay-area.

39. John Elflein，“1960-2020 年美国医疗保健支出占 GDP 的百分比”，Statista，2020 年 6 月 8 日，www.statista.com/statistics/184968/us-health-expenditure-as-percent-of-gdp-since-1960/。

39. John Elflein, “U.S. health care expenditure as a percentage of GDP 1960–2020,” Statista, June 8, 2020, www.statista.com/statistics/184968/us-health-expenditure-as-percent-of-gdp-since-1960/.

40. “医疗保健支出和融资”，OCED.stat，2020 年 5 月 15 日访问，stats.oecd.org/Index.aspx?DataSetCode=SHA。

40. “Healthcare expenditure and financing,” OCED.stat, accessed May 15, 2020, stats.oecd.org/Index.aspx?DataSetCode=SHA.

41. William J. Baumol 和 William G. Bowen，《表演艺术的经济困境：关于戏剧、歌剧、音乐和舞蹈常见问题的研究》，麻省理工学院出版社，1966 年。

41. William J. Baumol and William G. Bowen, Performing Arts, The Economic Dilemma: A Study of Problems Common to Theater, Opera, Music and Dance, MIT Press, 1966.

42. Michael Maiello，“诊断威廉·鲍莫尔的成本病”， 《芝加哥布斯评论》，2017 年 5 月 18 日，review.chicagobooth.edu/economics/2017/article/diagnosing-william-baumol-s-cost-disease 。

42. Michael Maiello, “Diagnosing William Baumol’s cost disease,” Chicago Booth Review, May 18, 2017, review.chicagobooth.edu/economics/2017/article/diagnosing-william-baumol-s-cost-disease.

43. “未来智能医院的 7 款医疗机器人”，Nanalyze，2020 年 4 月 6 日，www.nanalyze.com/2020/04/healthcare-robots-smart-hospital/。

43. “7 healthcare robots for the smart hospital of the future,” Nanalyze, April 6, 2020, www.nanalyze.com/2020/04/healthcare-robots-smart-hospital/.

44. Daphne Sashin，“机器人加入新斯坦福医院的员工队伍”， 《斯坦福医学新闻》 ，2019 年 11 月 4 日，med.stanford.edu/news/all-news/2019/11/robots-join - the-workforce-at-the-new-stanford-hospital -.html。

44. Daphne Sashin, “Robots join workforce at the new Stanford Hospital,” Stanford Medicine News, November 4, 2019, med.stanford.edu/news/all-news/2019/11/robots-join-the-workforce-at-the-new-stanford-hospital-.html.

45. Diego Ardila、Atilla P. Kiraly、Sujeeth Bharadwaj 等人，“通过低剂量胸部计算机断层扫描的三维深度学习进行端到端肺癌筛查”， 《自然医学》，第 25 卷，第 954–961 页 (2019)，2019 年 5 月 20 日，www.nature.com/articles/s41591-019-0447-x。

45. Diego Ardila, Atilla P. Kiraly, Sujeeth Bharadwaj et al., “End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography,” Nature Medicine, volume 25, pp. 954–961 (2019), May 20, 2019, www.nature.com/articles/s41591-019-0447-x.

46. Karen Hao，“医生正在使用人工智能对新冠肺炎患者进行分类。这些工具可能会继续存在”， 《麻省理工技术评论》 ，2020 年 4 月 23 日，www.technologyreview.com/2020/04 / 23/1000410/ai-triage-covid-19-patients -health-care。

46. Karen Hao, “Doctors are using AI to triage COVID-19 patients. The tools may be here to stay,” MIT Technology Review, April 23, 2020, www.technologyreview.com/2020/04/23/1000410/ai-triage-covid-19-patients-health-care.

47. Creative Distribution Lab，《Geoffrey Hinton：论放射学（视频）》，YouTube，2016 年 11 月 24 日，www.youtube.com/ watch ?reload=9&v=2HMPRXstSvQ。（2016 年机器学习和智能市场会议的一部分。）

47. Creative Distribution Lab, “Geoffrey Hinton: On radiology (video),” YouTube, November 24, 2016, www.youtube.com/watch?reload=9&v=2HMPRXstSvQ. (Part of the Machine Learning and the Market for Intelligence 2016 conference.)

48. Alex Bratt，“为什么放射科医生不必惧怕深度学习”， 《美国放射学会杂志》，第 16 卷，第 9期，A 部分，第 1190–1192 页（2019 年 9 月），2019 年 4 月 18 日，www.jacr.org/article/S1546-1440(19)30198-X/fulltext。

48. Alex Bratt, “Why radiologists have nothing to fear from deep learning,” Journal of the American College of Radiology, volume 16, issue 9, Part A, pp. 1190–1192 (September 2019), April 18, 2019, www.jacr.org/article/S1546-1440(19)30198-X/fulltext.

49. Ray Sipherd，“大多数医生不想让你知道的美国第三大死亡原因”，CNBC，2018 年 2 月 22 日，www.cnbc.com/2018/02/22/medical-errors-third-leading-cause-of-death-in-america.html。

49. Ray Sipherd, “The third-leading cause of death in US most doctors don’t want you to know about,” CNBC, February 22, 2018, www.cnbc.com/2018/02/22/medical-errors-third-leading-cause-of-death-in-america.html.

50. Elise Reuter，“研究表明，使用医疗 IT 初创公司的软件可以减少用药错误”， MedCity News ，2019 年 12 月 24 日，medcitynews.com/2019/12/study-shows-reduction - in-medication-errors-using-health-it-startups-software /。

50. Elise Reuter, “Study shows reduction in medication errors using health IT startup’s software,” MedCity News, December 24, 2019, medcitynews.com/2019/12/study-shows-reduction-in-medication-errors-using-health-it-startups-software/.

51. Adam Vaughan，“谷歌正在接管 DeepMind 的 NHS 合同——我们应该担心吗？” 《新科学家》 ，2019 年 9 月 27 日，www.newscientist.com/article/2217939-google-is-taking - over-deepminds-nhs-contracts-should-we-be-worried /。

51. Adam Vaughan, “Google is taking over DeepMind’s NHS contracts—should we be worried?,” New Scientist, September 27, 2019, www.newscientist.com/article/2217939-google-is-taking-over-deepminds-nhs-contracts-should-we-be-worried/.

52. Clive Thompson，“人工智能能帮助你吗？”， 《纽约时报》 ，2018 年 11 月 14 日，www.nytimes.com/interactive/2018/ 11 /14/magazine/tech-design-ai-chatbot.html。

52. Clive Thompson, “May A.I. help you?,” New York Times, November 14, 2018, www.nytimes.com/interactive/2018/11/14/magazine/tech-design-ai-chatbot.html.

53. Blair Hanley Frank，“Woebot 为其 AI 治疗师筹集 800 万美元”， VentureBeat ，2018 年 3 月 1 日，venturebeat.com/2018/03/01/woebot-raises-8 - million-for-its-ai-therapist /。

53. Blair Hanley Frank, “Woebot raises $8 million for its AI therapist,” VentureBeat, March 1, 2018, venturebeat.com/2018/03/01/woebot-raises-8-million-for-its-ai-therapist/.

54. Ariana Eunjung Cha，《沃森的下一个壮举？对抗癌症》，《 华盛顿邮报》，2015 年 6 月 27 日，www.washingtonpost.com/sf/national/2015/06/27/watsons-next -feat- taking-on-cancer /。

54. Ariana Eunjung Cha, “Watson’s next feat? Taking on cancer,” Washington Post, June 27, 2015, www.washingtonpost.com/sf/national/2015/06/27/watsons-next-feat-taking-on-cancer/.

55. Mary Chris Jaklevic，“MD 安德森癌症中心的 IBM Watson 项目失败，相关新闻也失败了”， 《健康新闻评论》 ，2017 年 2 月 23 日，www.healthnewsreview.org/2017/02/md - anderson-cancer-centers-ibm-watson-project-fails-journalism-related /。

55. Mary Chris Jaklevic, “MD Anderson Cancer Center’s IBM Watson project fails, and so did the journalism related to it,” Health News Review, February 23, 2017, www.healthnewsreview.org/2017/02/md-anderson-cancer-centers-ibm-watson-project-fails-journalism-related/.

56. Mark Anderson，“惊喜！2020 年不是自动驾驶汽车之年”， IEEE Spectrum ，2020 年 4 月 22 日，spectrum.ieee.org/transportation/self - driving/surprise-2020-is-not-the-year-for-selfdriving-cars。

56. Mark Anderson, “Surprise! 2020 is not the year for self-driving cars,” IEEE Spectrum, April 22, 2020, spectrum.ieee.org/transportation/self-driving/surprise-2020-is-not-the-year-for-selfdriving-cars.

57. Alex Knapp，“Aurora 首席执行官 Chris Urmson 表示，五年内将有数百辆自动驾驶汽车上路”， 《福布斯》，2019 年 10 月 29 日，www.forbes.com/sites/alexknapp/2019 / 10/29/aurora-ceo-chris-urmson-says-therell-be-hundreds-of-self-driving-cars-on-the-road-in-five-years /。

57. Alex Knapp, “Aurora CEO Chris Urmson says there’ll be hundreds of self-driving cars on the road in five years,” Forbes, October 29, 2019, www.forbes.com/sites/alexknapp/2019/10/29/aurora-ceo-chris-urmson-says-therell-be-hundreds-of-self-driving-cars-on-the-road-in-five-years/.

58. Lex Fridman，“Chris Urmson：Aurora、谷歌、卡内基梅隆大学和 DARPA 的自动驾驶汽车”，人工智能播客，第 28 集，2019 年 7 月 22 日，lexfridman.com/chris-urmson/。（提供视频和音频播客。）

58. Lex Fridman, “Chris Urmson: Self-driving cars at Aurora, Google, CMU, and DARPA,” Artificial Intelligence Podcast, episode 28, July 22, 2019, lexfridman.com/chris-urmson/. (Video and audio podcast available.)

59. Stefan Seltz-Axmacher，“Starsky Robotics 的终结”，Starsky Robotics 10-4 Labs 博客，2020 年 3 月 19 日，medium.com/starsky-robotics-blog/the-end-of-starsky-robotics-acb8a6a8a5f5。

59. Stefan Seltz-Axmacher, “The end of Starsky Robotics,” Starsky Robotics 10-4 Labs Blog, March 19, 2020, medium.com/starsky-robotics-blog/the-end-of-starsky-robotics-acb8a6a8a5f5.

60. Sam Dean，“得益于风险投资，Uber 车费很便宜。但免费出行时代结束了吗？” ， 《洛杉矶时报》 ，2019 年 5 月 11 日，www.latimes.com/business/technology/la- fi -tn-uber-ipo-lyft-fare-increase-20190511-story.html。

60. Sam Dean, “Uber fares are cheap, thanks to venture capital. But is that free ride ending?,” Los Angeles Times, May 11, 2019, www.latimes.com/business/technology/la-fi-tn-uber-ipo-lyft-fare-increase-20190511-story.html.

61. Darrell Etherington，“Waymo 现已在模拟中自动驾驶 100 亿英里”，TechCrunch，2019 年 7 月 10 日，techcrunch.com/2019/07/10/waymo-has-now-driven-10-billion-autonomous-miles-in-simulation/。

61. Darrell Etherington, “Waymo has now driven 10 billion autonomous miles in simulation,” TechCrunch, July 10, 2019, techcrunch.com/2019/07/10/waymo-has-now-driven-10-billion-autonomous-miles-in-simulation/.

62。Waymo网站，2020 年 5 月 20 日访问，waymo.com/。

62. Waymo website, accessed May 20, 2020, waymo.com/.

63. Ray Kurzweil，《加速回报定律》，Kurzweil 图书馆博客，2001 年 3 月 7 日，www.kurzweilai.net/the-law-of-accelerating-returns。

63. Ray Kurzweil, “The Law of Accelerating Returns,” Kurzweil Library Blog, March 7, 2001, www.kurzweilai.net/the-law-of-accelerating-returns.

64 .泰勒·科文， 《大停滞：美国如何尝到了现代史上所有唾手可得的果实，然后病倒了，并（最终）感觉好起来》，Dutton，2011 年。

64. Tyler Cowen, The Great Stagnation: How America Ate All the Low-Hanging Fruit of Modern History, Got Sick, and Will (Eventually) Feel Better, Dutton, 2011.

65.罗伯特·戈登， 《美国增长的兴衰：内战以来美国的生活水平》，普林斯顿大学出版社，2016年。

65. Robert J. Gordon, The Rise and Fall of American Growth: The U.S. Standard of Living Since the Civil War, Princeton University Press, 2016.

66 . Nicholas Bloom、Charles I. Jones、John Van Reenen 和 Michael Webb，“创意是否越来越难找到了？”《 美国经济评论》，第 110 卷，第 4 期，第 1104-1144 页（2020 年 4 月），www.aeaweb.org/articles?id=10.1257/aer.20180338，第 1138 页。

66. Nicholas Bloom, Charles I. Jones, John Van Reenen and Michael Webb, “Are ideas getting harder to find?” American Economic Review, volume 110, issue 4, pp. 1104–1144 (April 2020), www.aeaweb.org/articles?id=10.1257/aer.20180338, p. 1138.

67.同上，第1104页。

67. Ibid., p. 1104.

68.同上，第1104页。

68. Ibid., p. 1104.

69. Sam Lemonick，“探索化学空间：人工智能能否带我们前往人类从未涉足过的地方？” ， 《化学与工程新闻》 ，2020 年 4 月 6 日，cen.acs.org/physical-chemistry / computational-chemistry/Exploring-chemical-space-AI-take/98/i13。

69. Sam Lemonick, “Exploring chemical space: Can AI take us where no human has gone before?,” Chemical and Engineering News, April 6, 2020, cen.acs.org/physical-chemistry/computational-chemistry/Exploring-chemical-space-AI-take/98/i13.

70.同上。

70. Ibid.

71.代尔夫特理工大学，“研究人员利用人工智能设计新材料”，Phys.org，2019 年 10 月 14 日，phys.org/news/2019-10-material-artificial-intelligence.html。

71. Delft University of Technology, “Researchers design new material using artificial intelligence,” Phys.org, October 14, 2019, phys.org/news/2019-10-material-artificial-intelligence.html.

72. Beatrice Jin，“人工智能如何助力推动新材料发现”，康奈尔研究，2020 年 5 月 22 日访问，research.cornell.edu/research/how-ai-helps-advance-new-materials-discovery。

72. Beatrice Jin, “How AI helps to advance new materials discovery,” Cornell Research, accessed May 22, 2020, research.cornell.edu/research/how-ai-helps-advance-new-materials-discovery.

73. Savanna Hoover，“人工智能与材料科学相遇”，德克萨斯 A&M 大学工程新闻，2018 年 12 月 17 日，engineering.tamu.edu/news/2018/12/artificial-intelligence-meets-materials-science.html。

73. Savanna Hoover, “Artificial intelligence meets materials science,” Texas A&M University Engineering News, December 17, 2018, engineering.tamu.edu/news/2018/12/artificial-intelligence-meets-materials-science.html.

74. Kyle Wiggers，“Kebotix 筹集 1150 万美元，利用人工智能和机器人技术实现实验室实验自动化”， VentureBeat ，2020 年 4 月 16 日，venturebeat.com/2020/04/16/kebotix-raises-11-5-million-to - automate-lab-experiments-with-ai-and-robotics /。

74. Kyle Wiggers, “Kebotix raises $11.5 million to automate lab experiments with AI and robotics,” VentureBeat, April 16, 2020, venturebeat.com/2020/04/16/kebotix-raises-11-5-million-to-automate-lab-experiments-with-ai-and-robotics/.

75. Simon Smith，“230 家初创公司在药物研发中使用人工智能”，BenchSci 博客，2020 年 4 月 8 日更新，blog.benchsci.com/startups-using-artificial-intelligence-in-drug-discovery。

75. Simon Smith, “230 startups using artificial intelligence in drug discovery,” BenchSci Blog, updated April 8, 2020, blog.benchsci.com/startups-using-artificial-intelligence-in-drug-discovery.

76.福特，与达芙妮·科勒的访谈，载于 《情报建筑师》，第 388 页。

76. Ford, Interview with Daphne Koller, in Architects of Intelligence, p. 388.

77. Ned Pagliarulo，“GSK 的 Hal Barron 预测，人工智能对药物研发的影响将迅速显现”， BioPharma Dive ，2019 年 11 月 21 日，www.biopharmadive.com/news/gsk-hal-barron-ai -drug -discovery -prediction-daphne-koller/567855/。

77. Ned Pagliarulo, “AI’s impact in drug discovery is coming fast, predicts GSK’s Hal Barron,” BioPharma Dive, November 21, 2019, www.biopharmadive.com/news/gsk-hal-barron-ai-drug-discovery-prediction-daphne-koller/567855/.

78. Anne Trafton，“人工智能产生新抗生素”， 《麻省理工学院新闻》 ，2020 年 2 月 20 日，news.mit.edu/2020/artificial-intelligence-identifies-new- antibiotic -0220。

78. Anne Trafton, “Artificial intelligence yields new antibiotic,” MIT News, February 20, 2020, news.mit.edu/2020/artificial-intelligence-identifies-new-antibiotic-0220.

79. Richard Staines，“Exscientia 宣称人工智能创造的药物进入临床，为世界首创”， Pharmaphorum ，2020 年 1 月 30 日，pharmaphorum.com/news/exscientia-claims-world-first - as-ai-created-drug-enters-clinic /。

79. Richard Staines, “Exscientia claims world first as AI-created drug enters clinic,” Pharmaphorum, January 30, 2020, pharmaphorum.com/news/exscientia-claims-world-first-as-ai-created-drug-enters-clinic/.

80. Matt Reynolds，“DeepMind 的 AI 正越来越接近其第一个大型现实世界应用”， 《连线》，2020 年 1 月 15 日，www.wired.co.uk/article/deepmind-protein-folding-alphafold 。

80. Matt Reynolds, “DeepMind’s AI is getting closer to its first big real-world application,” Wired, January 15, 2020, www.wired.co.uk/article/deepmind-protein-folding-alphafold.

81。Semantic Scholar 网站，2020 年 5 月 25 日访问，pages.semanticscholar.org/about-us。

81. Semantic Scholar website, accessed May 25, 2020, pages.semanticscholar.org/about-us.

82.同上。

82. Ibid.

83. Khari Johnson，“微软、白宫和艾伦研究所向医学和 NLP 研究人员发布冠状病毒数据集”， VentureBeat ，2020 年 3 月 16 日，venturebeat.com/2020 / 03/16/microsoft-white-house-and-allen-institute-release-coronavirus-data-set-for-medical-and-nlp-researchers /。

83. Khari Johnson, “Microsoft, White House, and Allen Institute release coronavirus data set for medical and NLP researchers,” VentureBeat, March 16, 2020, venturebeat.com/2020/03/16/microsoft-white-house-and-allen-institute-release-coronavirus-data-set-for-medical-and-nlp-researchers/.

84. “CORD-19：COVID-19 开放研究数据集”，Semantic Scholar，2020 年 5 月 6 日访问，www.semanticscholar.org/ cord19 。

84. “CORD-19: COVID-19 Open Research Dataset,” Semantic Scholar, accessed May 6, 2020, www.semanticscholar.org/cord19.

第四章构建智能机器的探索

CHAPTER 4. THE QUEST TO BUILD INTELLIGENT MACHINES

1.塞缪尔·巴特勒，《机器中的达尔文，致编辑的一封信》，《新闻报》，新西兰基督城，1863 年 6 月 13 日。

1. Samuel Butler, “Darwin among the machines, a letter to the editors,” The Press, Christchurch, New Zealand, June 13, 1863.

2.阿兰·图灵，《计算机器和智能》， 《Mind 》，第LIX 卷，第 236 期，第 433-460 页（1950 年 10 月）。

2. Alan Turing, “Computing machinery and intelligence,” Mind, volume LIX, issue 236, pp. 433–460 (October 1950).

3. J. McCarthy、ML Minsky、N. Rochester 和 CE Shannon，《达特茅斯人工智能夏季研究项目提案》，1955 年 8 月 31 日，raysolomonoff.com/dartmouth/boxa/dart564props.pdf。

3. J. McCarthy, M. L. Minsky, N. Rochester and C. E. Shannon, “A proposal for the Dartmouth Summer Research Project on Artificial Intelligence,” August 31, 1955, raysolomonoff.com/dartmouth/boxa/dart564props.pdf.

4.布拉德·达拉赫，《遇见第一个电子人 Shaky：拥有自己思想的机器的迷人和可怕之处》， 《生活》，1970 年 11 月 20 日，第 58D 页。

4. Brad Darrach, “Meet Shaky, the first electronic person: The fascinating and fearsome reality of a machine with a mind of its own,” LIFE, November 20, 1970, p. 58D.

5.同上。

5. Ibid.

6. Warren McCulloch 和 Walter Pitts，《神经活动中内在观念的逻辑演算》，《数学生物物理学报》，第5 卷，第 4 期，第 115-133 页（1943 年 12 月）。

6. Warren McCulloch and Walter Pitts, “A logical calculus of ideas immanent in nervous activity,” Bulletin of Mathematical Biophysics, volume 5, issue 4, pp. 115–133 (December 1943).

7.马丁·福特，《雷·库兹韦尔访谈》， 《智能建筑师：来自人工智能建造者的真相》（Packt Publishing，2018 年，第 228 页）。

7. Martin Ford, Interview with Ray Kurzweil, in Architects of Intelligence: The Truth about AI from the People Building It, Packt Publishing, 2018, p. 228.

8. Marvin Minsky 和 Seymour Papert， 《感知器：计算几何导论》，麻省理工学院出版社，1969 年。

8. Marvin Minsky and Seymour Papert, Perceptrons: An Introduction to Computational Geometry, MIT Press, 1969.

9. Ford，采访 Yann LeCun，载于 《智能建筑师》，第 122 页。

9. Ford, Interview with Yann LeCun, in Architects of Intelligence, p. 122.

10. David E. Rumelhart、Geoffrey E. Hinton 和 Ronald J. Williams，《通过反向传播误差学习表示》，《 自然》 ，第323 卷，第 6088 期，第 533-536 页（1986 年），1986 年 10 月 9 日，www.nature.com/articles/323533a0。

10. David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams, “Learning representations by back-propagating errors,” Nature, volume 323, issue 6088, pp. 533–536 (1986), October 9, 1986, www.nature.com/articles/323533a0.

11.福特，与杰弗里·辛顿的访谈，载于 《智能建筑师》，第 73 页。

11. Ford, Interview with Geoffrey Hinton, in Architects of Intelligence, p. 73.

12. Dave Gershgorn，“改变人工智能研究乃至改变世界的数据”， Quartz ，2017 年 7 月 26 日，qz.com/1034972/the-data-that-changed-the-direction - of-ai-research-and-possibly-the-world/。

12. Dave Gershgorn, “The data that transformed AI research—and possibly the world,” Quartz, July 26, 2017, qz.com/1034972/the-data-that-changed-the-direction-of-ai-research-and-possibly-the-world/.

13.福特，与杰弗里·辛顿的访谈，载于 《智能建筑师》，第 77 页。

13. Ford, Interview with Geoffrey Hinton, in Architects of Intelligence, p. 77.

14.Jürgen Schmidhuber 发给 Martin Ford 的电子邮件，2019 年 1 月 28 日。

14. Email from Jürgen Schmidhuber to Martin Ford, January 28, 2019.

15. Jürgen Schmidhuber，“对‘深度学习阴谋’论文的批评（Nature 521 p 436）”，2015 年 6 月，people.idsia.ch/~juergen/deep-learning-conspiracy.html。

15. Jürgen Schmidhuber, “Critique of paper by ‘Deep Learning Conspiracy’ (Nature 521 p 436),” June 2015, people.idsia.ch/~juergen/deep-learning-conspiracy.html.

16. John Markoff，“当人工智能成熟时，它可能会称呼 Jürgen Schmidhuber 为‘爸爸’” ， 《纽约时报》，2016 年 11 月 27 日，www.nytimes.com/2016/11/27/technology/ artificial -intelligence -pioneer-jurgen-schmidhuber-overlooked.html。

16. John Markoff, “When A.I. matures, it may call Jürgen Schmidhuber ‘Dad,’” New York Times, November 27, 2016, www.nytimes.com/2016/11/27/technology/artificial-intelligence-pioneer-jurgen-schmidhuber-overlooked.html.

17. Robert Triggs，“成为‘AI 优先’公司对谷歌意味着什么”， Android Authority，2017 年 11 月 8 日，www.androidauthority.com/google-ai-first-812335/ 。

17. Robert Triggs, “What being an ‘AI first’ company means for Google,” Android Authority, November 8, 2017, www.androidauthority.com/google-ai-first-812335/.

18. Cade Metz，“谷歌的 AI 研究人员为何坐在老板旁边的办公桌上”， 《纽约时报》，2018 年 2 月 19 日，www.nytimes.com/2018/02/19/technology/ai-researchers-desks-boss.html 。

18. Cade Metz, “Why A.I. researchers at Google got desks next to the boss,” New York Times, February 19, 2018, www.nytimes.com/2018/02/19/technology/ai-researchers-desks-boss.html.

第五章深度学习与人工智能的未来

CHAPTER 5. DEEP LEARNING AND THE FUTURE OF ARTIFICIAL INTELLIGENCE

1. Martin Ford，采访 Geoffrey Hinton，载于《智能建筑师：来自人工智能建造者的真相》，Packt Publishing，2018 年，第 72-73 页。

1. Martin Ford, Interview with Geoffrey Hinton, in Architects of Intelligence: The Truth about AI from the People Building It, Packt Publishing, 2018, pp. 72–73.

2.Matt Reynolds，“新的计算机视觉挑战希望教会机器人以 3D 方式看世界”， 《新科学家》 ，2017 年 4 月 7 日，www.newscientist.com/article/2127131-new-computer-vision-challenge-wants-to- teach -robots-to-see-in-3d /。

2. Matt Reynolds, “New computer vision challenge wants to teach robots to see in 3D,” New Scientist, April 7, 2017, www.newscientist.com/article/2127131-new-computer-vision-challenge-wants-to-teach-robots-to-see-in-3d/.

3. Ashlee Vance，“硅谷最新的独角兽由一名 22 岁年轻人经营” ， 彭博商业周刊，2019 年 8 月 5 日，www.bloomberg.com/news/articles/2019-08-05 / scale-ai-is-silicon-valley-s-latest -unicorn。

3. Ashlee Vance, “Silicon Valley’s latest unicorn is run by a 22-year-old,” Bloomberg Businessweek, August 5, 2019, www.bloomberg.com/news/articles/2019-08-05/scale-ai-is-silicon-valley-s-latest-unicorn.

4. Volodymyr Mnih、Koray Kavukcuoglu、David Silver 等人。“通过深度强化学习玩 Atari，”DeepMind Research，2013 年 1 月 1 日，deepmind.com/research/publications/playing-atari-deep-reinforcement-learning。

4. Volodymyr Mnih, Koray Kavukcuoglu, David Silver et al. “Playing Atari with deep reinforcement learning,” DeepMind Research, January 1, 2013, deepmind.com/research/publications/playing-atari-deep-reinforcement-learning.

5. Volodymyr Mnih、Koray Kavukcuoglu、David Silver 等人，《通过深度强化学习实现人类层面的控制》， 《自然》，第518 卷，第 529-533 页（2015 年），2015 年 2 月 25 日，www.nature.com/articles/nature14236。

5. Volodymyr Mnih, Koray Kavukcuoglu, David Silver et al., “Human-level control through deep reinforcement learning,” Nature, volume 518, pp. 529–533 (2015), February 25, 2015, www.nature.com/articles/nature14236.

6.屠媛媛，《围棋：古老的智慧》， 《孔子学院杂志》第17卷，第46-51页（2011年11月），confuciusmag.com/go-game 。

6. Tu Yuanyuan, “The game of Go: Ancient wisdom,” Confucius Institute Magazine, volume 17, pp. 46–51 (November 2011), confuciusmag.com/go-game.

7. David Silver 和 Demis Hassabis，“AlphaGo：利用机器学习掌握古老的围棋游戏”，Google AI 博客，2016 年 1 月 27 日，ai.googleblog.com/2016/01/alphago-mastering-ancient-game-of-go.html。

7. David Silver and Demis Hassabis, “AlphaGo: Mastering the ancient game of Go with machine learning,” Google AI Blog, January 27, 2016, ai.googleblog.com/2016/01/alphago-mastering-ancient-game-of-go.html.

8. Matt Schiavenza，“中国的‘斯普特尼克时刻’与中美人工智能霸权之争”，亚洲协会博客，2018 年 9 月 25 日，asiasociety.org/blog/asia/chinas-sputnik-moment-and-sino-american-battle-ai-supremacy。

8. Matt Schiavenza, “China’s ‘Sputnik Moment’ and the Sino-American battle for AI supremacy,” Asia Society Blog, September 25, 2018, asiasociety.org/blog/asia/chinas-sputnik-moment-and-sino-american-battle-ai-supremacy.

9. John Markoff，“科学家看到深度学习程序的前景”， 纽约时报，2012 年 11 月 23 日，www.nytimes.com/2012/11 / 24/science/scientists-see-advances-in-deep-learning -a-part-of-artificial-intelligence.html。

9. John Markoff, “Scientists see promise in deep-learning programs,” New York Times, November 23, 2012, www.nytimes.com/2012/11/24/science/scientists-see-advances-in-deep-learning-a-part-of-artificial-intelligence.html.

10 . Dario Amodei 和 Danny Hernandez，“人工智能和计算”，OpenAI 博客，2018 年 5 月 16 日，openai.com/blog/ai-and-compute/。

10. Dario Amodei and Danny Hernandez, “AI and Compute,” OpenAI Blog, May 16, 2018, openai.com/blog/ai-and-compute/.

11. Will Knight，“Facebook 人工智能主管称该领域将很快‘碰壁’”， 《连线》，2019 年 12 月 4 日，www.wired.com/ story/facebooks-ai-says- field -hit-wall/。

11. Will Knight, “Facebook’s head of AI says the field will soon ‘hit the wall,’” Wired, December 4, 2019, www.wired.com/story/facebooks-ai-says-field-hit-wall/.

12. Kim Martineau，“缩小深度学习的碳足迹”， 《麻省理工学院新闻》，2020 年 8 月 7 日，news.mit.edu/2020/shrinking-deep-learning-carbon-footprint-0807 。

12. Kim Martineau, “Shrinking deep learning’s carbon footprint,” MIT News, August 7, 2020, news.mit.edu/2020/shrinking-deep-learning-carbon-footprint-0807.

13. “使用模式网络进行一般游戏”，Vicarious Research，2017 年 8 月 7 日，www.vicarious.com/2017/08/07/general-game-playing-with-schema-networks/。

13. “General game playing with schema networks,” Vicarious Research, August 7, 2017, www.vicarious.com/2017/08/07/general-game-playing-with-schema-networks/.

14. Sam Shead，“研究人员：我们是否正处于‘人工智能寒冬’的边缘？”，BBC 新闻，2020 年 1 月 12 日，www.bbc.com/news/technology-51064369。

14. Sam Shead, “Researchers: Are we on the cusp of an ‘AI winter’?,” BBC News, January 12, 2020, www.bbc.com/news/technology-51064369.

15. Filip Piekniewski，“人工智能寒冬即将来临”，Piekniewski 博客，2018 年 5 月 28 日，blog.piekniewski.info/2018/05/28/ai-winter-is-well-on-its-way/。

15. Filip Piekniewski, “AI winter is well on its way,” Piekniewski’s Blog, May 28, 2018, blog.piekniewski.info/2018/05/28/ai-winter-is-well-on-its-way/.

16 .福特，与杰弗里·迪恩的访谈，载于 《情报建筑师》，第 377 页。

16. Ford, Interview with Jeffery Dean, in Architects of Intelligence, p. 377.

17. Ford，与 Demis Hassabis 的访谈，载于 《智能建筑师》，第 171 页。

17. Ford, Interview with Demis Hassabis, in Architects of Intelligence, p. 171.

18. Andrea Banino、Caswell Barry、Dharshan Kumaran 和 Benigno Uria，“在人工智能代理中使用网格状表示进行导航”，DeepMind 研究博客，2018 年 5 月 9 日，deepmind.com/blog/article/grid-cells。

18. Andrea Banino, Caswell Barry, Dharshan Kumaran and Benigno Uria, “Navigating with grid-like representations in artificial agents,” DeepMind Research Blog, May 9, 2018, deepmind.com/blog/article/grid-cells.

19 . Ford，与 Demis Hassabis 的访谈，载于 《智能建筑师》，第 173 页。

19. Ford, Interview with Demis Hassabis, in Architects of Intelligence, p. 173.

20. Andrea Banino、Caswell Barry、Benigno Uria 等人，“在人工智能代理中使用网格状表示的基于矢量的导航”， 《自然》 ，第557 卷，第 429–433 页（2018 年），2018 年 5 月 9 日，www.nature.com/articles/s41586-018-0102-6。

20. Andrea Banino, Caswell Barry, Benigno Uria et al., “Vector-based navigation using grid-like representations in artificial agents,” Nature, volume 557, pp. 429–433 (2018), May 9, 2018, www.nature.com/articles/s41586-018-0102-6.

21. Will Dabney 和 Zeb Kurth-Nelson，“多巴胺和时间差异学习：神经科学与人工智能之间的富有成效的关系”，DeepMind 研究博客，2020 年 1 月 15 日，deepmind.com / blog/article/Dopamine-and-temporal-difference-learning-A-fruitful-relationship-between-neuroscience-and-AI。

21. Will Dabney and Zeb Kurth-Nelson, “Dopamine and temporal difference learning: A fruitful relationship between neuroscience and AI,” DeepMind Research Blog, January 15, 2020, deepmind.com/blog/article/Dopamine-and-temporal-difference-learning-A-fruitful-relationship-between-neuroscience-and-AI.

22. Tony Peng，《Yann LeCun 蛋糕类比 2.0》， Synced Review ，2019 年 2 月 22 日，medium.com/syncedreview/yann- lecun -cake-analogy-2-0-a361da560dae。

22. Tony Peng, “Yann LeCun Cake Analogy 2.0,” Synced Review, February 22, 2019, medium.com/syncedreview/yann-lecun-cake-analogy-2-0-a361da560dae.

23 . Ford，与 Demis Hassabis 的访谈，载于 《智能建筑师》，第 172-173 页。

23. Ford, Interview with Demis Hassabis, in Architects of Intelligence, pp. 172–173.

24. Jeremy Kahn，“人工智能在自然语言处理领域的突破对商业具有重大意义”， 《财富》，2020 年 1 月 20 日，fortune.com/2020/01/20/natural-language-processing- business /。

24. Jeremy Kahn, “A.I. breakthroughs in natural-language processing are big for business,” Fortune, January 20, 2020, fortune.com/2020/01/20/natural-language-processing-business/.

25 .福特，与大卫·费鲁奇的访谈，载于 《情报建筑师》，第 409 页。

25. Ford, Interview with David Ferrucci, in Architects of Intelligence, p. 409.

26.同上，第414页。

26. Ibid. p. 414.

27.《你信任这台电脑吗？》，2018 年 4 月 5 日发布，Papercut Films，doyoutrustthiscomputer.org/。

27. Do You Trust This Computer?, released April 5, 2018, Papercut Films, doyoutrustthiscomputer.org/.

28 . Ford，与 David Ferrucci 的访谈，载于 《情报建筑师》，第 414 页。

28. Ford, Interview with David Ferrucci, in Architects of Intelligence, p. 414.

29.雷·库兹韦尔， 《奇点临近：当人类超越生物学》，企鹅图书，2005年。

29. Ray Kurzweil, The Singularity Is Near: When Humans Transcend Biology, Penguin Books, 2005.

30.雷·库兹韦尔，《如何创造思维：揭秘人类思维》，企鹅图书，2012年。

30. Ray Kurzweil, How to Create a Mind: The Secret of Human Thought Revealed, Penguin Books, 2012.

31.福特，《雷·库兹韦尔访谈录》，载《智能建筑师》，第 230-231 页。

31. Ford, Interview with Ray Kurzweil, in Architects of Intelligence, pp. 230–231.

32. Mitch Kapor 和 Ray Kurzweil，“关于图灵测试的赌注：规则”，Kurzweil AI 博客，2002 年 4 月 9 日，www.kurzweilai.net/a-wager-on-the-turing-test-the-rules。

32. Mitch Kapor and Ray Kurzweil, “A wager on the Turing test: The rules,” Kurzweil AI Blog, April 9, 2002, www.kurzweilai.net/a-wager-on-the-turing-test-the-rules.

33. Sean Levinson，“谷歌高管为求长生不老每天服用 100 粒药丸”， 《精英日报》，2015 年 4 月 15 日，www.elitedaily.com/news/world/google-executive-taking-pills-live-forever/1001270 。

33. Sean Levinson, “A Google executive is taking 100 pills a day so he can live forever,” Elite Daily, April 15, 2015, www.elitedaily.com/news/world/google-executive-taking-pills-live-forever/1001270.

34.福特，《雷·库兹韦尔访谈录》，载《智能建筑师》，第 240-241 页。

34. Ford, Interview with Ray Kurzweil, in Architects of Intelligence, pp. 240–241.

35.同上，第230页。

35. Ibid., p. 230.

36.同上，第233页。

36. Ibid., p. 233.

37. Alec Radford、Jeffrey Wu、Dario Amodei 等人，“更好的语言模型及其含义”，OpenAI 博客，2019 年 2 月 14 日，openai.com/blog/better-language-models/。

37. Alec Radford, Jeffrey Wu, Dario Amodei et al., “Better language models and their implications,” OpenAI Blog, February 14, 2019, openai.com/blog/better-language-models/.

38. James Vincent，“OpenAI 的最新突破非常强大，但仍存在缺陷”， The Verge，2020 年 7 月 30 日，www.theverge.com/21346343/gpt-3-explainer-openai-examples-errors-agi-potential 。

38. James Vincent, “OpenAI’s latest breakthrough is astonishingly powerful, but still fighting its flaws,” The Verge, July 30, 2020, www.theverge.com/21346343/gpt-3-explainer-openai-examples-errors-agi-potential.

39. Gary Marcus 和 Ernest Davis，“GPT-3、Bloviator：OpenAI 的语言生成器不知道自己在说什么” ， 《麻省理工技术评论》 ，2020 年 8 月 22 日，www.technologyreview.com/2020/08 / 22/1007539/gpt3-openai-language-generator -artificial-intelligence-ai-opinion/。

39. Gary Marcus and Ernest Davis, “GPT-3, Bloviator: OpenAI’s language generator has no idea what it’s talking about,” MIT Technology Review, August 22, 2020, www.technologyreview.com/2020/08/22/1007539/gpt3-openai-language-generator-artificial-intelligence-ai-opinion/.

40.福特，与斯图尔特·罗素的访谈，载于《情报建筑师》，第 53 页。

40. Ford, Interview with Stuart Russell, in Architects of Intelligence, p. 53.

41. “OpenAI 创始人：短期 AGI 是一种严肃的可能性”， Synced，2018 年 11 月 13 日，syncedreview.com/2018/11/13/openai-founder-short-term-agi-is- a - serious-possibility/。

41. “OpenAI Founder: Short-Term AGI Is a Serious Possibility,” Synced, November 13, 2018, syncedreview.com/2018/11/13/openai-founder-short-term-agi-is-a-serious-possibility/.

42. Connie Loizos，“Sam Altman 与 StrictlyVC 对话（视频）”，YouTube，2019 年 5 月 18 日，youtu.be/TzcJlKg2Rc0，位置 39:00。

42. Connie Loizos, “Sam Altman in conversation with StrictlyVC (video),” YouTube, May 18, 2019, youtu.be/TzcJlKg2Rc0, location 39:00.

43. Luke Dormehl，“神经符号 AI 是人工智能的未来。它是这样运作的”， Digital Trends，2020 年 1 月 5 日，www.digitaltrends.com/cool-tech/neuro-symbolic-ai-the-future/ 。

43. Luke Dormehl, “Neuro-symbolic A.I. is the future of artificial intelligence. Here’s how it works,” Digital Trends, January 5, 2020, www.digitaltrends.com/cool-tech/neuro-symbolic-ai-the-future/.

44. Ford，与 Yoshua Bengio 的访谈，载于 《智能建筑师》，第 22 页。

44. Ford, Interview with Yoshua Bengio, in Architects of Intelligence, p. 22.

45. Ford，采访 Geoffrey Hinton，载于《智能建筑师》，第 84-85 页。

45. Ford, Interview with Geoffrey Hinton, in Architects of Intelligence, pp. 84–85.

46. Ford，采访 Yann LeCun，载于 《智能建筑师》，第 123 页。

46. Ford, Interview with Yann LeCun, in Architects of Intelligence, p. 123.

47. Anthony M. Zador，“对纯粹学习的批判以及人工神经网络可以从动物大脑中学到什么”， 《自然通讯》，第 10 卷，文章编号3770 (2019)，2019 年 8 月 21 日，www.nature.com/articles/s41467-019-11786-6。

47. Anthony M. Zador, “A critique of pure learning and what artificial neural networks can learn from animal brains,” Nature Communications, volume 10, article number 3770 (2019), August 21, 2019, www.nature.com/articles/s41467-019-11786-6.

48. Zoey Chong，“人工智能在斯坦福阅读理解测试中击败人类”，CNET，2018 年 1 月 16 日，www.cnet.com/news/new-results-show-ai-is-as-good-as- reading -comprehension-as-we-are/。

48. Zoey Chong, “AI beats humans in Stanford reading comprehension test,” CNET, January 16, 2018, www.cnet.com/news/new-results-show-ai-is-as-good-as-reading-comprehension-as-we-are/.

49.所有 Winograd 模式示例均取自：Ernest Davis，《Winograd 模式集合》，纽约大学计算机科学系，2011 年 9 月 8 日，cs.nyu.edu/davise/papers/WSOld.html。

49. All Winograd schema examples are taken from: Ernest Davis, “A collection of Winograd schemas,” New York University Department of Computer Science, September 8, 2011, cs.nyu.edu/davise/papers/WSOld.html.

50. Ford，与 Oren Etzioni 的访谈，载于《智能建筑师》，第 495-496 页。

50. Ford, Interview with Oren Etzioni, in Architects of Intelligence, pp. 495–496.

51.同上。

51. Ibid.

52. Ford，与 Yoshua Bengio 的访谈，载于 《智能建筑师》，第 21 页。

52. Ford, Interview with Yoshua Bengio, in Architects of Intelligence, p. 21.

53. Ford，采访 Yann LeCun，载于 《智能建筑师》，第 126-127 页。

53. Ford, Interview with Yann LeCun, in Architects of Intelligence, pp. 126–127.

54.同上，第130页。

54. Ibid., p. 130.

55. Ford，与 Judea Pearl 的访谈，载于《情报建筑师》，第 364 页。

55. Ford, Interview with Judea Pearl, in Architects of Intelligence, p. 364.

56. Ford，采访 Joshua Tenenbaum，载于 《情报建筑师》，第 471-472 页。

56. Ford, Interview with Joshua Tenenbaum, in Architects of Intelligence, pp. 471–472.

57. Ford，与 Judea Pearl 的访谈，载于《情报建筑师》，第 366 页。

57. Ford, Interview with Judea Pearl, in Architects of Intelligence, p. 366.

58. Will Knight，“人工智能先驱希望他的算法能够理解‘为什么’”， 《连线》，2019 年 10 月 8 日，www.wired.com/story/ai-pioneer-algorithms-understand-why/ 。

58. Will Knight, “An AI pioneer wants his algorithms to understand the ‘why,’” Wired, October 8, 2019, www.wired.com/story/ai-pioneer-algorithms-understand-why/.

59. Graham Allison， 《注定一战：中美能否逃脱修昔底德陷阱？》，霍顿·米夫林·哈考特出版社，2017 年。

59. Graham Allison, Destined for War: Can America and China Escape Thucydides’s Trap?, Houghton Mifflin Harcourt, 2017.

60. AlphaStar 团队，“AlphaStar：精通实时战略游戏 《星际争霸 II》 ”，DeepMind 研究博客，2019 年 1 月 24 日，deepmind.com/blog/article/alphastar-mastering-real - time-strategy-game-starcraft -ii。

60. The AlphaStar team, “AlphaStar: Mastering the real-time strategy game StarCraft II,” DeepMind Research Blog, January 24, 2019, deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii.

61 . Ford，与 Oren Etzioni 的访谈，载于 《情报建筑师》，第 494 页。

61. Ford, Interview with Oren Etzioni, in Architects of Intelligence, p. 494.

62.福特，《情报建筑师》，第528页。

62. Ford, Architects of Intelligence, p. 528.

63. “AI 时间线调查”，AI Impacts，2020 年 6 月 27 日访问，aiimpacts.org/ai-timeline-surveys/。

63. “AI timeline surveys,” AI Impacts, accessed June 27, 2020, aiimpacts.org/ai-timeline-surveys/.

第六章消失的工作岗位和人工智能的经济后果

CHAPTER 6. DISAPPEARING JOBS AND THE ECONOMIC CONSEQUENCES OF AI

1. David Axelrod，《Larry Summers》，《The Axe Files》（播客），第 98 集，2016 年 11 月 21 日，omny.fm/shows/the-axe-files - with-david-axelrod/ep-98-larry-summers。

1. David Axelrod, “Larry Summers,” The Axe Files (podcast), episode 98, November 21, 2016, omny.fm/shows/the-axe-files-with-david-axelrod/ep-98-larry-summers.

2. Sam Fleming 和 Brooke Fox，“投票支持特朗普的美国各州最容易受到工作自动化的影响”， 《金融时报》，2019 年 1 月 23 日，www.ft.com/content/cbf2a01e-1f41-11e9-b126-46fc3ad87c65 。

2. Sam Fleming and Brooke Fox, “US states that voted for Trump most vulnerable to job automation,” Financial Times, January 23, 2019, www.ft.com/content/cbf2a01e-1f41-11e9-b126-46fc3ad87c65.

3. Carol Graham，《了解绝望情绪在美国阿片类药物危机中的作用》，布鲁金斯学会，2019 年 10 月 15 日，www.brookings.edu/policy2020/votervital/how-can-policy-address-the-opioid-crisis-and-despair-in-america/。

3. Carol Graham, “Understanding the role of despair in America’s opioid crisis,” Brookings Institution, October 15, 2019, www.brookings.edu/policy2020/votervital/how-can-policy-address-the-opioid-crisis-and-despair-in-america/.

4.例如，请参阅：Carl Benedikt Frey 和 Michael A. Osborne，《就业的未来：工作对计算机化有多敏感？》，牛津马丁学院技术与就业项目，工作报告，2013 年 9 月 17 日，www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf，第 38 页。

4. See, for example: Carl Benedikt Frey and Michael A. Osborne, “The future of employment: How susceptible are jobs to computerisation?,” Oxford Martin School Programme on Technology and Employment, Working Paper, September 17, 2013, www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf, p. 38.

5.美国劳工统计局，“失业率（UNRATE）”，摘自圣路易斯联邦储备银行，2020 年 7 月 18 日，fred.stlouisfed.org/series/UNRATE；Greg Rosalsky，“我们是否接近充分就业？”，NPR Planet Money，2019 年 7 月 2 日，www.npr.org/sections/money/2019/07/02/737790095/are-we-even-close-to-full-employment。

5. U.S. Bureau of Labor Statistics, “Unemployment rate (UNRATE),” retrieved from Federal Reserve Bank of St. Louis, July 18, 2020, fred.stlouisfed.org/series/UNRATE; Greg Rosalsky, “Are we even close to full employment?,” NPR Planet Money, July 2, 2019, www.npr.org/sections/money/2019/07/02/737790095/are-we-even-close-to-full-employment.

6.经济合作与发展组织，“活动率：25-54 岁：美国男性（LRAC25MAUSM156S）”，摘自圣路易斯联邦储备银行，2020 年 7 月 17 日，fred.stlouisfed.org/series/LRAC25MAUSM156S。

6. Organization for Economic Co-operation and Development, “Activity rate: Aged 25–54: Males for the United States (LRAC25MAUSM156S),” retrieved from Federal Reserve Bank of St. Louis, July 17, 2020, fred.stlouisfed.org/series/LRAC25MAUSM156S.

7. “社会保障残疾保险趋势”，社会保障退休和残疾政策办公室，简报第 2019-01 号，2019 年 8 月，www.ssa.gov/policy/docs/briefing-papers/bp2019-01.html。

7. “Trends in Social Security Disability Insurance,” Social Security Office of Retirement and Disability Policy, Briefing Paper No. 2019-01, August 2019, www.ssa.gov/policy/docs/briefing-papers/bp2019-01.html.

8.美国劳工统计局，“劳动参与率（CIVPART）”，摘自圣路易斯联邦储备银行，2020 年 7 月 17 日，fred.stlouisfed.org/series/CIVPART。

8. U.S. Bureau of Labor Statistics, “Labor force participation rate (CIVPART),” retrieved from Federal Reserve Bank of St. Louis, July 17, 2020, fred.stlouisfed.org/series/CIVPART.

9.美国劳工统计局， “商业部门：所有人每小时实际产出（OPHPBS）”，摘自圣路易斯联邦储备银行，2020 年 7 月 22 日，fred.stlouisfed.org/series/OPHPBS；美国劳工统计局，“商业部门：每小时实际薪酬（PRS84006151）”，摘自圣路易斯联邦储备银行，2020 年 7 月 22 日，fred.stlouisfed.org/series/PRS84006151。

9. U.S. Bureau of Labor Statistics, “Business sector: Real output per hour of all persons (OPHPBS),” retrieved from Federal Reserve Bank of St. Louis, July 22, 2020, fred.stlouisfed.org/series/OPHPBS; U.S. Bureau of Labor Statistics, “Business sector: Real compensation per hour (PRS84006151),” retrieved from Federal Reserve Bank of St. Louis, July 22, 2020, fred.stlouisfed.org/series/PRS84006151.

10.世界银行，“美国基尼系数（SIPOVGINIUSA）”，摘自圣路易斯联邦储备银行，2020 年 7 月 20 日，fred.stlouisfed.org/series/SIPOVGINIUSA。

10. World Bank, “GINI index for the United States (SIPOVGINIUSA),” retrieved from Federal Reserve Bank of St. Louis, July 20, 2020, fred.stlouisfed.org/series/SIPOVGINIUSA.

11. Martha Ross 和 Nicole Bateman，“低薪工作比你想象的更加普遍，而且‘好工作’也不够多”，布鲁金斯学会，2019 年 11 月 21 日，www.brookings.edu/blog/the-avenue/2019/11/21/low-wage-work-is-more-pervasive-than-you-think-and-there-arent-enough-good-jobs-to-go-around/。

11. Martha Ross and Nicole Bateman, “Low-wage work is more pervasive than you think, and there aren’t enough ‘good jobs’ to go around,” Brookings Institution, November 21, 2019, www.brookings.edu/blog/the-avenue/2019/11/21/low-wage-work-is-more-pervasive-than-you-think-and-there-arent-enough-good-jobs-to-go-around/.

12. “美国私营部门就业质量指数（JQI）”，2020 年 7 月 15 日访问，www.jobqualityindex.com/。

12. “The U.S. Private Sector Job Quality Index (JQI),” accessed July 15, 2020, www.jobqualityindex.com/.

13. Gwynn Guilford，“美国劳动力大悖论：工作机会充足，但大多都是坏事”， Quartz ，2019 年 11 月 21 日，qz.com/1752676/the-job-quality-index - is-the-economic-indicator-weve-been-missing/。

13. Gwynn Guilford, “The great American labor paradox: Plentiful jobs, most of them bad,” Quartz, November 21, 2019, qz.com/1752676/the-job-quality-index-is-the-economic-indicator-weve-been-missing/.

14 . Elizabeth Redden，“41% 的应届毕业生从事不需要学位的工作”， 《Inside Higher Ed》，2020 年 2 月 18 日，www.insidehighered.com/quicktakes/2020/02/18/41-recent-grads-work-jobs-not-requiring-degree。

14. Elizabeth Redden, “41% of recent grads work in jobs not requiring a degree,” Inside Higher Ed, February 18, 2020, www.insidehighered.com/quicktakes/2020/02/18/41-recent-grads-work-jobs-not-requiring-degree.

15. “菲利普斯曲线可能被彻底打破”， 《经济学人》，2017 年 11 月 1 日，www.economist.com/graphic-detail/2017/11/01/the-phillips- curve - may-be-broken-for-good。

15. “The Phillips curve may be broken for good,” The Economist, November 1, 2017, www.economist.com/graphic-detail/2017/11/01/the-phillips-curve-may-be-broken-for-good.

16. Jeff Jeffrey，“美国公司现金滚滚，但他们越来越不敢花掉这些钱” ， 《商业期刊》 ，2018 年 12 月 12 日，www.bizjournals.com/bizjournals/news/2018/12 /12 /us-companies-are-hoarding-cash-and-theyre-growing.html。

16. Jeff Jeffrey, “U.S. companies are rolling in cash, and they’re growing increasingly fearful to spend it,” The Business Journals, December 12, 2018, www.bizjournals.com/bizjournals/news/2018/12/12/u-s-companies-are-hoarding-cash-and-theyre-growing.html.

17. Martin Ford， 《机器人的崛起：技术与未来失业的威胁》，Basic Books，2015 年，第 206-212 页。

17. Martin Ford, Rise of the Robots: Technology and the Threat of a Jobless Future, Basic Books, 2015, pp. 206–212.

18. Martin Ford，James Manyika 访谈，载于《智能建筑师：来自人工智能建造者的真相》，Packt Publishing，2018 年，第 285–286 页。

18. Martin Ford, Interview with James Manyika, in Architects of Intelligence: The Truth about AI from the People Building It, Packt Publishing, 2018, pp. 285–286.

19 . Nir Jaimovich 和 Henry E. Siu，“就业两极分化和无就业复苏”，美国国家经济研究局，工作文件 18334，2012 年 8 月发布，2018 年 11 月修订，www.nber.org/papers/w18334。

19. Nir Jaimovich and Henry E. Siu, “Job polarization and jobless recoveries,” National Bureau of Economic Research, Working Paper 18334, issued in August 2012, revised in November 2018, www.nber.org/papers/w18334.

20. Jacob Bunge 和 Jesse Newman，“受新冠疫情影响，泰森食品启用机器人屠夫” ， 《华尔街日报》 ，2020 年 7 月 9 日，www.wsj.com/articles/meatpackers-covid - safety-automation-robots-coronavirus -11594303535。

20. Jacob Bunge and Jesse Newman, “Tyson turns to robot butchers, spurred by coronavirus outbreaks,” Wall Street Journal, July 9, 2020, www.wsj.com/articles/meatpackers-covid-safety-automation-robots-coronavirus-11594303535.

21. Miso Robotics，“White Castle 选择 Miso Robotics 开启快餐行业人工智能新时代”，新闻稿，2020 年 7 月 14 日，www.prnewswire.com/news-releases/white-castle-selects-miso-robotics-for-a-new-era-of-artificial-intelligence-in-the-fast-food-industry-301092746.html。

21. Miso Robotics, “White Castle selects Miso Robotics for a new era of artificial intelligence in the fast food industry,” Press Release Newswire, July 14, 2020, www.prnewswire.com/news-releases/white-castle-selects-miso-robotics-for-a-new-era-of-artificial-intelligence-in-the-fast-food-industry-301092746.html.

22. James Manyika、Susan Lund、Michael Chui 等人，“就业岗位减少，就业岗位增加：未来工作对就业、技能和工资意味着什么”，麦肯锡全球研究院，2017 年 11 月 28 日，www.mckinsey.com/featured-insights/future-of-work/jobs-lost-jobs-gained-what-the-future-of-work-will-mean-for-jobs-skills-and-wages。

22. James Manyika, Susan Lund, Michael Chui, et al., “Jobs lost, jobs gained: What the future of work will mean for jobs, skills, and wages,” McKinsey Global Institute, November 28, 2017, www.mckinsey.com/featured-insights/future-of-work/jobs-lost-jobs-gained-what-the-future-of-work-will-mean-for-jobs-skills-and-wages.

23. Ferris Jabr，《Cache cab：出租车司机的大脑进化以适应伦敦街道》， 《科学美国人》，2011 年 12 月 8 日，www.scientificamerican.com/article/london-taxi-memory/ 。

23. Ferris Jabr, “Cache cab: Taxi drivers’ brains grow to navigate London’s streets,” Scientific American, December 8, 2011, www.scientificamerican.com/article/london-taxi-memory/.

24 . Kate Conger，“Facebook 开始规划永久性远程工作者”， 《纽约时报》，2020 年 5 月 21 日，www.nytimes.com/2020/05/21/technology/facebook-remote-work-coronavirus.html。

24. Kate Conger, “Facebook starts planning for permanent remote workers,” New York Times, May 21, 2020, www.nytimes.com/2020/05/21/technology/facebook-remote-work-coronavirus.html.

25. Alexandre Tanzi，“美国小企业前景黯淡，52% 预测将破产”， 彭博社，2020 年 5 月 6 日，www.bloomberg.com/news/articles/2020-05-06/majority - of-us-small-businesses -expect-to-close-survey-says。

25. Alexandre Tanzi, “Gloom grips U.S. small businesses, with 52% predicting failure,” Bloomberg, May 6, 2020, www.bloomberg.com/news/articles/2020-05-06/majority-of-u-s-small-businesses-expect-to-close-survey-says.

26 . Alfred Liu，“研究称，未来十年机器人将导致美国银行业岗位减少 20 万个” ， 彭博社，2019 年 10 月 1 日，www.bloomberg.com/news/articles/2019-10-02/robots-to-cut-200-000-us-bank-jobs-in-next-decade-study-says。

26. Alfred Liu, “Robots to cut 200,000 U.S. bank jobs in next decade, study says,” Bloomberg, October 1, 2019, www.bloomberg.com/news/articles/2019-10-02/robots-to-cut-200-000-u-s-bank-jobs-in-next-decade-study-says.

27.杰克·凯利，“人工智能正在取代高薪的华尔街工作”， 《福布斯》 ，2019 年 12 月 10 日，www.forbes.com/sites/jackkelly/2019/12/10/ artificial -intelligence-is-superseding-well-paying-wall-street-jobs /。

27. Jack Kelly, “Artificial intelligence is superseding well-paying Wall Street jobs,” Forbes, December 10, 2019, www.forbes.com/sites/jackkelly/2019/12/10/artificial-intelligence-is-superseding-well-paying-wall-street-jobs/.

28. “顶级医疗聊天机器人初创公司”，Tracxn，2020 年 10 月 20 日，tracxn.com/d/trending-themes/Startups-in-Healthcare-Chatbots。

28. “Top healthcare chatbots startups,” Tracxn, October 20, 2020, tracxn.com/d/trending-themes/Startups-in-Healthcare-Chatbots.

29. Celeste Barnaby、Satish Chandra 和 Frank Luan，“Aroma：使用机器学习进行代码推荐”，Facebook AI 博客，2019 年 4 月 4 日，ai.facebook.com/blog/aroma-ml-for-code-recommendation/。

29. Celeste Barnaby, Satish Chandra and Frank Luan, “Aroma: Using machine learning for code recommendation,” Facebook AI Blog, April 4, 2019, ai.facebook.com/blog/aroma-ml-for-code-recommendation/.

30. Will Douglas Heaven，“OpenAI 的新语言生成器 GPT-3 出奇地好——而且完全没有思维”， 《麻省理工技术评论》 ，2020 年 7 月 20 日，www.technologyreview.com/2020/07 / 20/1005454/openai-machine-learning-language-generator-gpt-3-nlp/。

30. Will Douglas Heaven, “OpenAI’s new language generator GPT-3 is shockingly good—and completely mindless,” MIT Technology Review, July 20, 2020, www.technologyreview.com/2020/07/20/1005454/openai-machine-learning-language-generator-gpt-3-nlp/.

31. Jacques Bughin、Jeongmin Seong、James Manyika 等人，“人工智能前沿笔记：模拟人工智能对世界经济的影响”，麦肯锡全球研究院，讨论文件，2018 年 9 月，www.mckinsey.com/~/media/McKinsey/Featured%20Insights/Artificial%20Intelligence/Notes%20from%20the%20frontier%20Modeling%20the%20impact%20of%20AI%20on%20the%20world%20economy/MGI-Notes-from-the-AI-frontier-Modeling-the-impact-of-AI-on-the-world-economy-September-2018.ashx。

31. Jacques Bughin, Jeongmin Seong, James Manyika, et al., “Notes from the AI frontier: Modeling the impact of AI on the world economy,” McKinsey Global Institute, Discussion Paper, September 2018, www.mckinsey.com/~/media/McKinsey/Featured%20Insights/Artificial%20Intelligence/Notes%20from%20the%20frontier%20Modeling%20the%20impact%20of%20AI%20on%20the%20world%20economy/MGI-Notes-from-the-AI-frontier-Modeling-the-impact-of-AI-on-the-world-economy-September-2018.ashx.

32. Anand S. Rao 和 Gerard Verweij，“确定奖项规模：人工智能对您的业务的真正价值是什么？您如何利用它？”，普华永道，2018 年 10 月，www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf。

32. Anand S. Rao and Gerard Verweij, “Sizing the prize: What’s the real value of AI for your business and how can you capitalise?,” PwC, October 2018, www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf.

33. Bughin 等人，“人工智能前沿笔记：模拟人工智能对世界经济的影响”，第 3 页。

33. Bughin et al., “Notes from the AI frontier: Modeling the impact of AI on the world economy,” p 3.

第七章中国与人工智能监控国家的崛起

CHAPTER 7. CHINA AND THE RISE OF THE AI SURVEILLANCE STATE

1.储百亮 (Chris Buckley)、莫祖尔 (Paul Mozur) 和奥斯汀·拉姆齐 (Austin Ramzy)，《中国如何将一座城市变成监狱》， 《纽约时报》，2019 年 4 月 4 日，www.nytimes.com/interactive/2019/04/04/world/asia/xinjiang-china-surveillance-prison.html。

1. Chris Buckley, Paul Mozur and Austin Ramzy, “How China turned a city into a prison,” New York Times, April 4, 2019, www.nytimes.com/interactive/2019/04/04/world/asia/xinjiang-china-surveillance-prison.html.

2. James Vincent，“中国网民发现习近平主席书架上有人工智能书籍”， The Verge ，2018 年 1 月 3 日，www.theverge.com/2018/1/3/16844364/ china -ai-xi-jinping-new-years-speech-books。

2. James Vincent, “Chinese netizens spot AI books on president Xi Jinping’s bookshelf,” The Verge, January 3, 2018, www.theverge.com/2018/1/3/16844364/china-ai-xi-jinping-new-years-speech-books.

3. Tom Simonite，“中国在人工智能研究领域正快速赶超美国”， 《连线》，2019 年 3 月 13 日，www.wired.com/story/china-catching-up-us-in-ai-research/ 。

3. Tom Simonite, “China is catching up to the US in AI research—fast,” Wired, March 13, 2019, www.wired.com/story/china-catching-up-us-in-ai-research/.

4. Robust Vision Challenge 网站，2020 年 7 月 25 日访问，www.robustvision.net / rvc2018.php。

4. Robust Vision Challenge website, accessed July 25, 2020, www.robustvision.net/rvc2018.php.

5.国防科技大学网站，2020年7月25日访问，english.nudt.edu.cn/About/index.htm。

5. National University of Defense Technology website, accessed July 25, 2020, english.nudt.edu.cn/About/index.htm.

6. Nicolas Thompson 和 Ian Bremmer，“威胁我们所有人的人工智能冷战”， 《连线》 ，2018 年 10 月 23 日，www.wired.com/story/ai-cold-war - china-could-doom-us-all/。

6. Nicolas Thompson and Ian Bremmer, “The AI Cold War that threatens us all,” Wired, October 23, 2018, www.wired.com/story/ai-cold-war-china-could-doom-us-all/.

7. Alex Hern，“中国审查谷歌 AlphaGo 与世界最优秀围棋选手的比赛”， 《卫报》 ，2017 年 5 月 24 日，www.theguardian.com/technology/2017/may/ 24/ china-censored-googles-alphago-match-against-worlds-best-go-player 。

7. Alex Hern, “China censored Google’s AlphaGo match against world’s best Go player,” The Guardian, May 24, 2017, www.theguardian.com/technology/2017/may/24/china-censored-googles-alphago-match-against-worlds-best-go-player.

8.中国国务院，《新一代人工智能发展规划》，中国国务院于 2017 年 7 月 20 日发布，由 Graham Webster、Rogier Creemers、Paul Triolo 和 Elsa Kania 翻译，新美国基金会，2017 年 8 月 1 日，www.newamerica.org/cybersecurity-initiative/digichina/blog/full-translation-chinas-new-generation-artificial-intelligence-development-plan-2017/。（中国政府文件原文：www.gov.cn/zhengce/content/2017-07/20/content_5211996.htm。）

8. China’s State Council, “New Generation Artificial Intelligence Development Plan,” issued by China’s State Council on July 20, 2017, translated by Graham Webster, Rogier Creemers, Paul Triolo and Elsa Kania, New America Foundation, August 1, 2017, www.newamerica.org/cybersecurity-initiative/digichina/blog/full-translation-chinas-new-generation-artificial-intelligence-development-plan-2017/. (Original Chinese government document: www.gov.cn/zhengce/content/2017-07/20/content_5211996.htm.)

9 . Lai Lin Thomala，“2008-2020 年中国互联网用户数量”，Statista，2020 年 4 月 30 日，www.statista.com/statistics/265140/number-of-internet-users-in-china/。

9. Lai Lin Thomala, “Number of internet users in China 2008–2020,” Statista, April 30, 2020, www.statista.com/statistics/265140/number-of-internet-users-in-china/.

10 . Lai Lin Thomala，“2008-2020 年中国互联网用户渗透率”，Statista，2020 年 4 月 30 日，www.statista.com/statistics/236963/penetration-rate-of-internet-users-in-china/。

10. Lai Lin Thomala, “Penetration rate of internet users in China 2008–2020,” Statista, April 30, 2020, www.statista.com/statistics/236963/penetration-rate-of-internet-users-in-china/.

11. Rachel Metz，“百度若采取完全谷歌式的举措，可在自动驾驶汽车领域击败谷歌” ， 《麻省理工技术评论》，2018 年 1 月 8 日，www.technologyreview.com/2018 / 01/08/146351/baidu-could-beat-google-in-self-driving-cars-with-a-totally-google-move /。

11. Rachel Metz, “Baidu could beat Google in self-driving cars with a totally Google move,” MIT Technology Review, January 8, 2018, www.technologyreview.com/2018/01/08/146351/baidu-could-beat-google-in-self-driving-cars-with-a-totally-google-move/.

12. Jon Russell，“前微软高管、著名人工智能专家陆奇加入百度担任首席运营官”，TechCrunch，2017 年 1 月 17 日，techcrunch.com/2017/01/16/qi-lu-joins-baidu-as-coo/。

12. Jon Russell, “Former Microsoft executive and noted AI expert Qi Lu joins Baidu as COO,” TechCrunch, January 17, 2017, techcrunch.com/2017/01/16/qi-lu-joins-baidu-as-coo/.

13. Martin Ford，采访 Demis Hassabis，载于《智能建筑师：来自人工智能建造者的真相》，Packt Publishing，2018 年，第 179 页。

13. Martin Ford, Interview with Demis Hassabis, in Architects of Intelligence: The Truth about AI from the People Building It, Packt Publishing, 2018, p. 179.

14. Field Cady 和 Oren Etzioni，“中国可能在人工智能研究领域超越美国”，艾伦人工智能研究所博客，2019 年 3 月 13 日，medium.com/ai2-blog/china-to-overtake-us-in-ai-research-8b6b1fe30595。

14. Field Cady and Oren Etzioni, “China may overtake US in AI research,” Allen Institute for AI Blog, March 13, 2019, medium.com/ai2-blog/china-to-overtake-us-in-ai-research-8b6b1fe30595.

15. Jeffrey Ding，“解读中国的人工智能梦：中国引领世界人工智能战略的背景、要素、能力和后果”，牛津大学人类未来研究所，2018 年 3 月，www.fhi.ox.ac.uk/wp-content/uploads/Deciphering_Chinas_AI-Dream.pdf。

15. Jeffrey Ding, “Deciphering China’s AI dream: The context, components, capabilities, and consequences of China’s strategy to lead the world in AI,” Future of Humanity Institute, University of Oxford, March 2018, www.fhi.ox.ac.uk/wp-content/uploads/Deciphering_Chinas_AI-Dream.pdf.

16 . Jeffrey Ding，“中国目前在人工智能领域的能力、政策和产业生态系统：在美中经济与安全审查委员会技术、贸易和军民融合听证会上的证词：中国对人工智能、新材料和新能源的追求”，2019 年 6 月 7 日，www.uscc.gov/sites/default/files/June%207%20Hearing_Panel%201_Jeffrey%20Ding_China%27s%20Current%20Capabilities%2C%20Policies%2C%20and%20Industrial%20Ecosystem%20in%20AI.pdf。

16. Jeffrey Ding, “China’s current capabilities, policies, and industrial ecosystem in AI: Testimony before the U.S.-China Economic and Security Review Commission Hearing on Technology, Trade, and Military-Civil Fusion: China’s Pursuit of Artificial Intelligence, New Materials, and New Energy,” June 7, 2019, www.uscc.gov/sites/default/files/June%207%20Hearing_Panel%201_Jeffrey%20Ding_China%27s%20Current%20Capabilities%2C%20Policies%2C%20and%20Industrial%20Ecosystem%20in%20AI.pdf.

17.李开复，“中国在人工智能方面能教给美国什么”， 纽约时报，2018 年 9 月 22 日，www.nytimes.com/2018/09/22/opinion/sunday/ai-china-united-states.html 。

17. Kai-Fu Lee, “What China can teach the U.S. about artificial intelligence,” New York Times, September 22, 2018, www.nytimes.com/2018/09/22/opinion/sunday/ai-china-united-states.html.

18. Kathrin Hille 和 Richard Waters，“中国‘军民融合’令华盛顿不安” ， 《金融时报》，2018 年 11 月 7 日，www.ft.com/content/8dcb534c-dbaf-11e8-9f04-38d397e6661c 。

18. Kathrin Hille and Richard Waters, “Washington unnerved by China’s ‘military-civil fusion,’” Financial Times, November 7, 2018, www.ft.com/content/8dcb534c-dbaf-11e8-9f04-38d397e6661c.

19. Scott Shane 和 Daisuke Wakabayashi，“‘战争生意’：谷歌员工抗议为五角大楼工作”， 《纽约时报》，2018 年 4 月 4 日，www.nytimes.com/2018/04 / 04/technology/google-letter-ceo-pentagon-project.html。

19. Scott Shane and Daisuke Wakabayashi, “‘The Business of War’: Google employees protest work for the Pentagon,” New York Times, April 4, 2018, www.nytimes.com/2018/04/04/technology/google-letter-ceo-pentagon-project.html.

20. Tom Simonite，《中国人脸识别巨头崛起的背后》， 《连线》，2019 年 9 月 3 日，www.wired.com/story/behind-rise-chinas-facial-recognition-giants/ 。

20. Tom Simonite, “Behind the rise of China’s facial-recognition giants,” Wired, September 3, 2019, www.wired.com/story/behind-rise-chinas-facial-recognition-giants/.

21. Paul Mozur 和 Aaron Krolik，“监控网络覆盖中国城市，赋予警方巨大权力”， 《纽约时报》，2019 年 12 月 17 日，www.nytimes.com/2019/12/17/technology/china-surveillance.html 。

21. Paul Mozur and Aaron Krolik, “A surveillance net blankets China’s cities, giving police vast powers,” New York Times, December 17, 2019, www.nytimes.com/2019/12/17/technology/china-surveillance.html.

22. Amy B. Wang，“一名嫌疑人试图混入 6 万名音乐会观众中，被中国的面部识别摄像头抓住”， 《华盛顿邮报》 ，2018 年 4 月 13 日，www.washingtonpost.com/news/worldviews / wp/2018/04/13/china-crime-facial-recognition -cameras-catch-suspect-at-concert-with-60000-people/。

22. Amy B. Wang, “A suspect tried to blend in with 60,000 concertgoers. China’s facial-recognition cameras caught him,” Washington Post, April 13, 2018, www.washingtonpost.com/news/worldviews/wp/2018/04/13/china-crime-facial-recognition-cameras-catch-suspect-at-concert-with-60000-people/.

23.保罗·莫祖尔，“揭秘中国反乌托邦梦想：人工智能、耻辱和大量摄像头” ， 《纽约时报》，2018 年 7 月 8 日，nytimes.com/2018/07/08/business/china-surveillance-technology.html 。

23. Paul Mozur, “Inside China’s dystopian dreams: A.I., shame and lots of cameras,” New York Times, July 8, 2018, nytimes.com/2018/07/08/business/china-surveillance-technology.html.

24. Paul Moser，“一个月 50 万次人脸扫描：中国如何利用人工智能来分析少数群体”， 《纽约时报》 ，2019 年 4 月 14 日，www.nytimes.com/2019/04 / 14/technology/china-surveillance-artificial-intelligence-racial-profiling.html。

24. Paul Moser, “One month, 500,000 face scans: How China is using A.I. to profile a minority,” New York Times, April 14, 2019, www.nytimes.com/2019/04/14/technology/china-surveillance-artificial-intelligence-racial-profiling.html.

25.同上。

25. Ibid.

26. Simina Mistreanu，《中国社会信用实验室内的生活》， 《外交政策》，2018 年 4 月 3 日，foreignpolicy.com/2018/04/03/life-inside - chinas-social-credit-laboratory/。

26. Simina Mistreanu, “Life inside China’s social credit laboratory,” Foreign Policy, April 3, 2018, foreignpolicy.com/2018/04/03/life-inside-chinas-social-credit-laboratory/.

27. Echo Huang，“中国违反垃圾分类规定者将面临被评为垃圾信用评级的风险”， Quartz ，2018 年 1 月 8 日，qz.com/1173975/garbage-sorting-violators - in-china-risk-getting-a-junk-credit-rating /。

27. Echo Huang, “Garbage-sorting violators in China now risk being punished with a junk credit rating,” Quartz, January 8, 2018, qz.com/1173975/garbage-sorting-violators-in-china-risk-getting-a-junk-credit-rating/.

28. Maya Wang，《中国令人心寒的‘社会信用’黑名单》，人权观察，2017 年 12 月 12 日，www.hrw.org/news/2017/12/13/chinas-chilling-social-credit-blacklist。

28. Maya Wang, “China’s chilling ‘social credit’ blacklist,” Human Rights Watch, December 12, 2017, www.hrw.org/news/2017/12/13/chinas-chilling-social-credit-blacklist.

29. Nicole Kobie，“中国社会信用体系的复杂真相”， 《连线》，2019 年 6 月 7 日，www.wired.co.uk/article/china-social-credit-system-explained 。

29. Nicole Kobie, “The complicated truth about China’s social credit system,” Wired, June 7, 2019, www.wired.co.uk/article/china-social-credit-system-explained.

30. Steven Feldstein，《人工智能监控的全球扩张》，卡内基国际和平基金会，2019 年 9 月 17 日，carnegieendowment.org/2019/09/17/global-expansion-of-ai-surveillance-pub-79847。

30. Steven Feldstein, “The global expansion of AI surveillance,” Carnegie Endowment for International Peace, September 17, 2019, carnegieendowment.org/2019/09/17/global-expansion-of-ai-surveillance-pub-79847.

31. Yuan Yang 和 Madhumita Murgia，“面部识别：中国如何垄断监控市场”， 《金融时报》，2019 年 12 月 6 日，www.ft.com/content/6f1a8f48-1813-11ea-9ee4-11f260415385 。

31. Yuan Yang and Madhumita Murgia, “Facial recognition: How China cornered the surveillance market,” Financial Times, December 6, 2019, www.ft.com/content/6f1a8f48-1813-11ea-9ee4-11f260415385.

32. Russell Brandon，“针对华为的案件解析”， The Verge，2019 年 5 月 22 日，www.theverge.com/2019/5/22/18634401/huawei-ban-trump-case-infrastructure-fears-google-microsoft-arm-security 。

32. Russell Brandon, “The case against Huawei, explained,” The Verge, May 22, 2019, www.theverge.com/2019/5/22/18634401/huawei-ban-trump-case-infrastructure-fears-google-microsoft-arm-security.

33. Will Knight，“特朗普针对中国的最新攻击瞄准了人工智能公司”， 《连线》，2019 年 10 月 9 日，www.wired.com/ story/trumps-salvo-against - china-targets-ai-firms/。

33. Will Knight, “Trump’s latest salvo against China targets AI firms,” Wired, October 9, 2019, www.wired.com/story/trumps-salvo-against-china-targets-ai-firms/.

34. Kashmir Hill，“这家神秘的公司可能会终结我们所知的隐私”， 《纽约时报》 ，2020 年 1 月 18 日，www.nytimes.com/2020/01/18/technology/clearview-privacy- facial -recognition.html。

34. Kashmir Hill, “The secretive company that might end privacy as we know it,” New York Times, January 18, 2020, www.nytimes.com/2020/01/18/technology/clearview-privacy-facial-recognition.html.

35.同上。

35. Ibid.

36.同上。

36. Ibid.

37. Ryan Mac、Caroline Haskins 和 Logan McDonald，“Clearview 的面部识别应用程序已被司法部、ICE、梅西百货、沃尔玛和 NBA 使用”， BuzzFeed新闻，2020 年 2 月 27 日，www.buzzfeednews.com/article/ryanmac/clearview-ai-fbi-ice-global-law-enforcement。

37. Ryan Mac, Caroline Haskins and Logan McDonald, “Clearview’s facial recognition app has been used by the Justice Department, ICE, Macy’s, Walmart, and the NBA,” BuzzFeed News, February 27, 2020, www.buzzfeednews.com/article/ryanmac/clearview-ai-fbi-ice-global-law-enforcement.

38. Alfred Ng 和 Steven Musil，“Clearview AI 因收集面部识别数据遭谷歌和 Facebook 勒令停止运营”，CNET，2020 年 2 月 5 日，www.cnet.com/news/clearview-ai-hit-with-cease-and-desist-from-google-over-facial-recognition-collection/。

38. Alfred Ng and Steven Musil, “Clearview AI hit with cease-and-desist from Google, Facebook over facial recognition collection,” CNET, February 5, 2020, www.cnet.com/news/clearview-ai-hit-with-cease-and-desist-from-google-over-facial-recognition-collection/.

39. Zack Whittaker，“Apple 已封锁 Clearview AI 的 iPhone 应用程序，原因是其违反了其规定”， TechCrunch，2020 年 2 月 28 日，techcrunch.com/2020/02/28/apple-ban-clearview-iphone/ 。

39. Zack Whittaker, “Apple has blocked Clearview AI’s iPhone app for violating its rules,” TechCrunch, February 28, 2020, techcrunch.com/2020/02/28/apple-ban-clearview-iphone/.

40. Nick Statt，“美国公民自由联盟起诉面部识别公司 Clearview AI，称其为隐私的‘噩梦’”， The Verge ，2020 年 5 月 28 日，www.theverge.com/2020 / 5/28/21273388/aclu-clearview-ai-lawsuit -facial-recognition-database-illinois-biometric-laws。

40. Nick Statt, “ACLU sues facial recognition firm Clearview AI, calling it a ‘nightmare scenario’ for privacy,” The Verge, May 28, 2020, www.theverge.com/2020/5/28/21273388/aclu-clearview-ai-lawsuit-facial-recognition-database-illinois-biometric-laws.

41. Paul Bischoff，“监控摄像头统计数据：哪些城市的闭路电视摄像头最多？”，Comparitech，2019 年 8 月 1 日，www.comparitech.com/vpn-privacy/the-worlds-most-surveilled-cities/。

41. Paul Bischoff, “Surveillance camera statistics: Which cities have the most CCTV cameras?,” Comparitech, August 1, 2019, www.comparitech.com/vpn-privacy/the-worlds-most-surveilled-cities/.

42. “伦敦警察厅将部署面部识别摄像头”，BBC，2020 年 1 月 30 日，www.bbc.com/news/uk-51237665。

42. “Met Police to deploy facial recognition cameras,” BBC, January 30, 2020, www.bbc.com/news/uk-51237665.

43. Clare Garvie、Alvaro Bedoya和Jonathan Frankle，《永久排队：美国不受监管的警察人脸识别》，乔治城大学隐私与技术法律中心，2016 年 10 月 18 日，www.perpetuallineup.org/。

43. Clare Garvie, Alvaro Bedoya and Jonathan Frankle, “The perpetual line-up: Unregulated police face recognition in America,” Georgetown Law Center on Privacy and Technology, October 18, 2016, www.perpetuallineup.org/.

44. “大都会警察局将部署面部识别摄像头。 ”

44. “Met Police to deploy facial recognition cameras.”

45. London Real，“Jonathan Haidt——自由放养的孩子：如何给予孩子更多自由（视频）”，2018 年 10 月 27 日，www.youtube.com/watch?v=GPTei2sroIk。

45. London Real, “Jonathan Haidt—Free range kids: How to give your children more freedom (video),” October 27, 2018, www.youtube.com/watch?v=GPTei2sroIk.

46. Isabella Garcia，“面部识别能否克服种族偏见？”， 《Yes!》杂志，2020 年 4 月 16 日，www.yesmagazine.org/social-justice/2020/ 04 /16privacy-facial-recognition/。

46. Isabella Garcia, “Can facial recognition overcome its racial bias?,” Yes! Magazine, April 16, 2020, www.yesmagazine.org/social-justice/2020/04/16privacy-facial-recognition/.

47. Sasha Ingber，“面部识别软件错误地将 28 名立法者识别为犯罪嫌疑人”，美国国家公共电台，2018 年 7 月 26 日，www.npr.org/2018/07 / 26/632724239/facial-recognition-software-wrongly-identifies-28-lawmakers-as-crime-suspects。

47. Sasha Ingber, “Facial recognition software wrongly identifies 28 lawmakers as crime suspects,” NPR, July 26, 2018, www.npr.org/2018/07/26/632724239/facial-recognition-software-wrongly-identifies-28-lawmakers-as-crime-suspects.

48. Patrick Grother、Mei Ngan 和 Kayee Hanaoka，“人脸识别供应商测试 (FRVT) 第 3 部分：人口统计影响”，美国国家标准与技术研究所，2019 年 12 月，nvlpubs.nist.gov/nistpubs/ir/2019/NIST.IR.8280.pdf。

48. Patrick Grother, Mei Ngan and Kayee Hanaoka, “Face Recognition Vendor Test (FRVT) Part 3: Demographic effects,” National Institute of Standards and Technology, December 2019, nvlpubs.nist.gov/nistpubs/ir/2019/NIST.IR.8280.pdf.

49.加西亚，“面部识别能否克服种族偏见？”

49. Garcia, “Can facial recognition overcome its racial bias?”

50. Amy Hawkins，“北京的‘老大哥’科技需要非洲面孔”， 《外交政策》 ，2018 年 7 月 24 日，foreignpolicy.com/2018/07/24/beijings-big - brother-tech-needs-african-faces /。

50. Amy Hawkins, “Beijing’s big brother tech needs African faces,” Foreign Policy, July 24, 2018, foreignpolicy.com/2018/07/24/beijings-big-brother-tech-needs-african-faces/.

第八章人工智能风险

CHAPTER 8. AI RISKS

1. “虚假声音‘帮助网络犯罪分子窃取现金’”，BBC 新闻，2019 年 7 月 8 日，www.bbc.com/news/technology-48908736。

1. “Fake voices ‘help cyber-crooks steal cash,’” BBC News, July 8, 2019, www.bbc.com/news/technology-48908736.

2. Martin Giles，《GAN 之父：赋予机器想象力的人》， 《麻省理工技术评论》，2018 年 2 月 21 日，www.technologyreview.com/2018/02/21/145289/the-ganfather-the-man-whos- given -machines-the-gift-of-imagination /。

2. Martin Giles, “The GANfather: The man who’s given machines the gift of imagination,” MIT Technology Review, February 21, 2018, www.technologyreview.com/2018/02/21/145289/the-ganfather-the-man-whos-given-machines-the-gift-of-imagination/.

3. James Vincent，“观看 Jordan Peele 使用 AI 让巴拉克·奥巴马发表有关假新闻的公共服务公告”， The Verge ，2018 年 4 月 17 日，www.theverge.com/tldr/2018/4/17/17247334/ai- fake -news-video-barack-obama-jordanpeele-buzzfeed。

3. James Vincent, “Watch Jordan Peele use AI to make Barack Obama deliver a PSA about fake news,” The Verge, April 17, 2018, www.theverge.com/tldr/2018/4/17/17247334/ai-fake-news-video-barack-obama-jordanpeele-buzzfeed.

4. Sensity，“2019 年深度伪造现状：格局、威胁和影响”，2019 年 9 月，sensity.ai/reports/。

4. Sensity, “The state of deepfakes 2019: Landscape, threats, and impact,” September 2019, sensity.ai/reports/.

5. Ian Sample，“什么是深度伪造——如何发现它们？” 《卫报》 ，2020 年 1 月 13 日，www.theguardian.com/technology/2020 / jan/13/what-are-deepfakes-and-how-can-you-spot -them。

5. Ian Sample, “What are deepfakes—and how can you spot them?,” The Guardian, January 13, 2020, www.theguardian.com/technology/2020/jan/13/what-are-deepfakes-and-how-can-you-spot-them.

6. Lex Fridman，《Ian Goodfellow：生成对抗网络 (GAN)》，人工智能播客，第 19 集，2019 年 4 月 18 日，lexfridman.com/ian-goodfellow/。（提供视频和音频播客。）

6. Lex Fridman, “Ian Goodfellow: Generative Adversarial Networks (GANs),” Artificial Intelligence Podcast, episode 19, April 18, 2019, lexfridman.com/ian-goodfellow/. (Video and audio podcast available.)

7. JJ McCorvey，“这家图像认证初创公司正在打击虚假社交媒体账户、篡改照片、深度伪造等”， Fast Company，2019 年 2 月 19 日，www.fastcompany.com/90299000/truepic-most-innovative-companies-2019 。

7. J.J. McCorvey, “This image-authentication startup is combating faux social media accounts, doctored photos, deep fakes, and more,” Fast Company, February 19, 2019, www.fastcompany.com/90299000/truepic-most-innovative-companies-2019.

8. Ian Goodfellow、Nicolas Papernot、Sandy Huang 等人，“利用对抗性示例攻击机器学习”，OpenAI 博客，2017 年 2 月 24 日，openai.com/blog/adversarial-example-research/。

8. Ian Goodfellow, Nicolas Papernot, Sandy Huang, et al., “Attacking machine learning with adversarial examples,” OpenAI Blog, February 24, 2017, openai.com/blog/adversarial-example-research/.

9. Anant Jain，“通过对抗性攻击破坏神经网络”，Towards Data Science，2019 年 2 月 9 日，towardsdatascience.com/breaking-neural-networks-with-adversarial-attacks-f4290a9a45aa。

9. Anant Jain, “Breaking neural networks with adversarial attacks,” Towards Data Science, February 9, 2019, towardsdatascience.com/breaking-neural-networks-with-adversarial-attacks-f4290a9a45aa.

10.同上。

10. Ibid.

11. 《Slaughterbots》，2017 年 11 月 12 日发布，Space Digital，www.youtube.com/watch?reload=9&v=9CO6M2HsoIA 。

11. Slaughterbots, released November 12, 2017, Space Digital, www.youtube.com/watch?reload=9&v=9CO6M2HsoIA.

12. Stuart Russell，“制造致命的自主武器比制造自动驾驶汽车更容易。必须签署一项新条约”， 《安全时报》，2018 年 2 月，www.the-security-times.com/ building-a-lethal-autonomous-weapon-is - easier-than-building-a-self-driving-car-a-new-treaty-is-necessary/。

12. Stuart Russell, “Building a lethal autonomous weapon is easier than building a self-driving car. A new treaty is necessary,” The Security Times, February 2018, www.the-security-times.com/building-a-lethal-autonomous-weapon-is-easier-than-building-a-self-driving-car-a-new-treaty-is-necessary/.

13.马丁·福特，斯图尔特·拉塞尔访谈，载于 《智能建筑师：来自人工智能建造者的真相》，Packt Publishing，2018 年，第 59 页。

13. Martin Ford, Interview with Stuart Russell, in Architects of Intelligence: The Truth about AI from the People Building It, Packt Publishing, 2018, p. 59.

14. “各国对杀手机器人的看法”，停止杀手机器人运动，2019 年 8 月 21 日，www.stopkillerrobots.org/wp-content/uploads/2019/08/KRC_CountryViews21Aug2019.pdf。

14. “Country views on killer robots,” Campaign to Stop Killer Robots, August 21, 2019, www.stopkillerrobots.org/wp-content/uploads/2019/08/KRC_CountryViews21Aug2019.pdf.

15. “俄罗斯、美国试图使杀手机器人合法化”，“阻止杀手机器人运动”，2019 年 8 月 22 日，www.stopkillerrobots.org/2019/08/russia-united-states-attempt-to-legitimize-killer-robots/。

15. “Russia, United States attempt to legitimize killer robots,” Campaign to Stop Killer Robots, August 22, 2019, www.stopkillerrobots.org/2019/08/russia-united-states-attempt-to-legitimize-killer-robots/.

16. Zachary Kallenborn，“大规模杀伤性无人机群：将武装全自动无人机群宣布为大规模杀伤性武器的理由”，现代战争研究所，2020 年 5 月 28 日，mwi.usma.edu/swarms-mass-destruction-case-declaring-armed-fully-autonomous-drone-swarms-wmd/。

16. Zachary Kallenborn, “Swarms of mass destruction: The case for declaring armed and fully autonomous drone swarms as WMD,” Modern War Institute, May 28, 2020, mwi.usma.edu/swarms-mass-destruction-case-declaring-armed-fully-autonomous-drone-swarms-wmd/.

17. Kris Osborn，“陆军新版 10 吨机器人来了”， 《国家利益》，2020 年 5 月 21 日，nationalinterest.org/blog/buzz/here-come-armys-new-class-10-ton-robots-156351 。

17. Kris Osborn, “Here come the Army’s new class of 10-ton robots,” National Interest, May 21, 2020, nationalinterest.org/blog/buzz/here-come-armys-new-class-10-ton-robots-156351.

18.Rachel England，“美国空军正准备进行人机与人工智能的空中缠斗”， Engadget ，2020 年 6 月 8 日，www.engadget.com/the-air-force-will-pit -an-autonomous -fighter-drone-against-a-pilot-121526011.html。

18. Rachel England, “The US Air Force is preparing a human versus AI dogfight,” Engadget, June 8, 2020, www.engadget.com/the-air-force-will-pit-an-autonomous-fighter-drone-against-a-pilot-121526011.html.

19. Kris Osborn，“机器人大战？现在中国拥有半自主地面作战机器人”， 《国家利益》 ，2020 年 6 月 15 日，nationalinterest.org/blog/buzz/robot-vs-robot-war - now-china-has-semi-autonomous-fighting-ground-robots -162782。

19. Kris Osborn, “Robot vs. robot war? Now China has semi-autonomous fighting ground robots,” National Interest, June 15, 2020, nationalinterest.org/blog/buzz/robot-vs-robot-war-now-china-has-semi-autonomous-fighting-ground-robots-162782.

20. Neil Johnson、Guannan Zhao、Eric Hunsader 等人，《超越人类反应时间的新机器生态的突然崛起》，《 自然科学报告》，第 3 卷，文章编号 2627（2013年），2013 年 9 月 11 日，www.nature.com/articles/srep02627。

20. Neil Johnson, Guannan Zhao, Eric Hunsader, et al., “Abrupt rise of new machine ecology beyond human response time,” Nature Scientific Reports, volume 3, article number 2627 (2013), September 11, 2013, www.nature.com/articles/srep02627.

21.福特，与斯图尔特·罗素的访谈，载于《情报建筑师》，第 59 页。

21. Ford, Interview with Stuart Russell, in Architects of Intelligence, p. 59.

22. Jeffrey Dastin，“亚马逊放弃秘密 AI 招聘工具，该工具显示出对女性的偏见”，路透社，2018 年 10 月 10 日，www.reuters.com/article/us-amazon-com- jobs- automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G。

22. Jeffrey Dastin, “Amazon scraps secret AI recruiting tool that showed bias against women,” Reuters, October 10, 2018, www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G.

23. Julia Angwin、Jeff Larson、Surya Mattu 和 Lauren Kirchner，《机器偏见》， Propublica，2016 年 5 月 23 日，www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing 。

23. Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner, “Machine bias,” Propublica, May 23, 2016, www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing.

24.同上。

24. Ibid.

25 .福特，与詹姆斯·曼尼卡的访谈，载于 《情报建筑师》，第 279 页。

25. Ford, Interview with James Manyika, in Architects of Intelligence, p. 279.

26 . Ford，《李飞飞访谈录》，载 《智能建筑师》，第 157 页。

26. Ford, Interview with Fei-Fei Li, in Architects of Intelligence, p. 157.

27.史蒂芬·霍金、斯图尔特·罗素、马克斯·泰格马克和弗兰克·威尔切克，《史蒂芬·霍金：《超越》着眼于人工智能的含义——但我们是否足够重视人工智能？》《 独立报》，2014 年 5 月 1 日，www.independent.co.uk/news/science/stephen-hawking-transcendence-looks-at-the-implications-of-artificial-intelligence-but-are-we-taking-ai-seriously-enough-9313474.html。

27. Stephen Hawking, Stuart Russell, Max Tegmark and Frank Wilczek, “Stephen Hawking: ‘Transcendence looks at the implications of artificial intelligence—but are we taking AI seriously enough?,’” The Independent, May 1, 2014, www.independent.co.uk/news/science/stephen-hawking-transcendence-looks-at-the-implications-of-artificial-intelligence-but-are-we-taking-ai-seriously-enough-9313474.html.

28. Nick Bostrom， 《超级智能：路径、危险、策略》，牛津大学出版社，2014年，第vii页。

28. Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, Oxford University Press, 2014, p. vii.

29. Matt McFarland，“伊隆·马斯克：‘我们利用人工智能召唤恶魔’” ， 《华盛顿邮报》 ，2014 年 10 月 24 日，www.washingtonpost.com/news/innovations/wp/2014 / 10/24/elon-musk-with-artificial-intelligence-we-are-summoning-the-demon /。

29. Matt McFarland, “Elon Musk: ‘With artificial intelligence we are summoning the demon,’” Washington Post, October 24, 2014, www.washingtonpost.com/news/innovations/wp/2014/10/24/elon-musk-with-artificial-intelligence-we-are-summoning-the-demon/.

30. Sam Harris，“我们能否在不失去对人工智能的控制的情况下制造出它？ (视频)”， TED 演讲，2016 年 6 月，www.ted.com/talks/sam_harris_can_we_build_ai_without_losing_control_over_it?language=en。

30. Sam Harris, “Can we build AI without losing control over it? (video),” TED Talk, June 2016, www.ted.com/talks/sam_harris_can_we_build_ai_without_losing_control_over_it?language=en.

31. Irving John Good，《关于第一台超智能机器的推测》，《 计算机高级》，第 6 卷，第 31-88 页（1965 年），vtechworks.lib.vt.edu/bitstream/handle/10919/89424/TechReport05-3.pdf 。

31. Irving John Good, “Speculations concerning the first ultraintelligent machine,” Advanced in Computers, volume 6, pp. 31–88 (1965), vtechworks.lib.vt.edu/bitstream/handle/10919/89424/TechReport05-3.pdf.

32. Jesselyn Cook，“数百人分享掉入 YouTube 推荐兔子洞的故事”， 《赫芬顿邮报》，2019 年 10 月 15 日，www.huffpost.com/entry/youtube-recommendation-rabbit-hole-mozilla_n_5da5c470e4b08f3654912991 。

32. Jesselyn Cook, “Hundreds of people share stories about falling down YouTube’s recommendation rabbit hole,” Huffington Post, October 15, 2019, www.huffpost.com/entry/youtube-recommendation-rabbit-hole-mozilla_n_5da5c470e4b08f3654912991.

33. Stuart Russell，《 与人类兼容：人工智能与控制问题》，Viking，2019 年，第 173-177 页。

33. Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control, Viking, 2019, pp. 173–177.

34. Stuart Russell，“如何在超人人工智能阻止我们之前阻止它”， 《纽约时报》，2019 年 10 月 8 日，www.nytimes.com/2019/10/08/opinion/artificial-intelligence.html 。

34. Stuart Russell, “How to stop superhuman A.I. before it stops us,” New York Times, October 8, 2019, www.nytimes.com/2019/10/08/opinion/artificial-intelligence.html.

35.福特，采访罗德尼·布鲁克斯，载于《情报建筑师》，第 440-441 页。

35. Ford, Interview with Rodney Brooks, in Architects of Intelligence, pp. 440–441.

三十五

结论：两种人工智能未来

CONCLUSION: TWO AI FUTURES

1. Rebecca Heilweil，“大型科技公司不再向警方出售面部识别技术”， Recode ，2020 年 6 月 11 日，www.vox.com/recode/2020/6/10/21287194/amazon - microsoft-ibm-facial-recognition -moratorium-police。

1. Rebecca Heilweil, “Big tech companies back away from selling facial recognition to police,” Recode, June 11, 2020, www.vox.com/recode/2020/6/10/21287194/amazon-microsoft-ibm-facial-recognition-moratorium-police.

2. Joseph Zeballos-Roig，“卡马拉·哈里斯支持每月 2,000 美元的刺激支票，以帮助美国人摆脱疫情带来的困境——而且她长期支持民主党向人们提供更多资金的计划”，《商业内幕》，2020 年 8 月 15 日，www.businessinsider.com/kamala-harris-biden-monthly-stimulus-checks-economic-policy-support-vice-2020-8。

2. Joseph Zeballos-Roig, “Kamala Harris supports $2,000 monthly stimulus checks to help Americans claw out of pandemic ruin—and she’s long backed plans for Democrats to give people more money,” Business Insider, August 15, 2020, www.businessinsider.com/kamala-harris-biden-monthly-stimulus-checks-economic-policy-support-vice-2020-8.

3. Bob Berwyn，“‘应对气候变化 12 年’（现在是 11 年）到底意味着什么？” ， 《气候新闻内幕》，2019 年 8 月 27 日，insideclimatenews.org/news/27082019/12 - years -climate-change-explained-ipcc-science-solutions。

3. Bob Berwyn, “What does ‘12 years to act on climate change’ (now 11 years) really mean?,” Inside Climate News, August 27, 2019, insideclimatenews.org/news/27082019/12-years-climate-change-explained-ipcc-science-solutions.

4.比尔·盖茨，《新冠疫情很可怕。气候变化可能会更糟》，《盖茨笔记》，2020 年 8 月 4 日，www.gatesnotes.com/Energy/Climate-and-COVID-19。

4. Bill Gates, “COVID-19 is awful. Climate change could be worse,” Gates Notes, August 4, 2020, www.gatesnotes.com/Energy/Climate-and-COVID-19.

5.比尔·盖茨，《气候变化和 75% 的问题》，盖茨笔记，2018 年 10 月 17 日，www.gatesnotes.com/Energy/My-plan-for-fighting-climate-change。

5. Bill Gates, “Climate change and the 75% problem,” Gates Notes, October 17, 2018, www.gatesnotes.com/Energy/My-plan-for-fighting-climate-change.

6. Nicholas Bloom、Charles I. Jones、John Van Reenen 和 Michael Webb，“创意是否越来越难找到了？” 《美国经济评论》，第 110 卷，第 4期，第 1104-1144 页（2020 年 4 月），www.aeaweb.org/articles?id=10.1257/aer.20180338，第 1138 页。

6. Nicholas Bloom, Charles I. Jones, John Van Reenen and Michael Webb, “Are ideas getting harder to find?,” American Economic Review, volume 110, issue 4, pp. 1104–1144 (April 2020), www.aeaweb.org/articles?id=10.1257/aer.20180338, p. 1138.

7. Mark Aguiar、Mark Bils、Kerwin Kofi Charles 和 Erik Hurst，《休闲奢侈品与年轻男性的劳动力供给》，美国国家经济研究局，工作报告 23552，2017 年 6 月，www.nber.org/papers/w23552。

7. Mark Aguiar, Mark Bils, Kerwin Kofi Charles and Erik Hurst, “Leisure luxuries and the labor supply of young men,” National Bureau of Economic Research, Working Paper 23552, June 2017, www.nber.org/papers/w23552.

赵晓晓

XIAOXIAO ZHAO

马丁·福特是一位未来学家，著有《机器人的崛起》等多部，荣膺《金融时报》年度商业图书奖。此外，他还著有《智能建筑师》和《隧道中的灯光》。他还是一家硅谷软件开发公司的创始人。他关于人工智能对社会影响的 TED 演讲已被观看超过 300 万次，他的文章见诸《纽约时报》、《财富》、《福布斯》、《大西洋月刊》、《华盛顿邮报》、《哈佛商业评论》、《卫报》和《金融时报》。福特是一位广受欢迎的演讲者，也是人工智能领域的顶尖专家。他现居加利福尼亚州桑尼维尔。

MARTIN FORD is a futurist and the author of the New York Times bestseller Rise of the Robots, which won the Financial Times Business Book of the Year Award; Architects of Intelligence; and The Lights in the Tunnel. He is also the founder of a Silicon Valley–based software development firm. His TED Talk on the impact of artificial intelligence on society has been viewed over 3 million times, and his writing has appeared in the New York Times, Fortune, Forbes, the Atlantic, the Washington Post, Harvard Business Review, the Guardian, and the Financial Times. Ford is a sought-after speaker and a leading expert on artificial intelligence. He lives in Sunnyvale, California.

推特：@MFordFuture

Twitter: @MFordFuture

马丁·福特

ALSO BY MARTIN FORD

隧道中的灯光：

自动化、加速技术

和未来经济

(Acculant Publishing，2009)

The Lights in the Tunnel:

Automation, Accelerating Technology

and the Economy of the Future

(Acculant Publishing, 2009)

机器人的崛起：

技术与未来失业的威胁

(Basic Books，2015)

Rise of the Robots:

Technology and the Threat of a Jobless Future

(Basic Books, 2015)

《智能建筑师：

人工智能建造者讲述的真相》

（Packt Publishing，2018 年）

Architects of Intelligence:

The Truth about AI from the People Building It

(Packt Publishing, 2018)

脚注

Footnote

脚注

Footnote

脚注

Footnote

脚注

Footnote

第一章 新兴颠覆

CHAPTER 1. THE EMERGING DISRUPTION

第 2 章 人工智能作为新电力

CHAPTER 2. AI AS THE NEW ELECTRICITY

第 3 章 超越炒作

CHAPTER 3. BEYOND HYPE

第四章 构建智能机器的探索

CHAPTER 4. THE QUEST TO BUILD INTELLIGENT MACHINES

第五章 深度学习与人工智能的未来

CHAPTER 5. DEEP LEARNING AND THE FUTURE OF ARTIFICIAL INTELLIGENCE

第六章 消失的工作岗位和人工智能的经济后果

CHAPTER 6. DISAPPEARING JOBS AND THE ECONOMIC CONSEQUENCES OF AI

第七章 中国与人工智能监控国家的崛起

CHAPTER 7. CHINA AND THE RISE OF THE AI SURVEILLANCE STATE

第八章 人工智能风险

CHAPTER 8. AI RISKS

结论：两种人工智能未来

CONCLUSION: TWO AI FUTURES

第一章新兴颠覆

第 2 章人工智能作为新电力

第 3 章超越炒作

第四章构建智能机器的探索

第五章深度学习与人工智能的未来

第六章消失的工作岗位和人工智能的经济后果

第七章中国与人工智能监控国家的崛起

第八章人工智能风险